mirror of
https://github.com/openai/whisper.git
synced 2025-09-13 11:10:11 +00:00
Compare commits
No commits in common. "main" and "v20230117" have entirely different histories.
3
.gitattributes
vendored
3
.gitattributes
vendored
@ -1,3 +0,0 @@
|
||||
# Override jupyter in Github language stats for more accurate estimate of repo code languages
|
||||
# reference: https://github.com/github/linguist/blob/master/docs/overrides.md#generated-code
|
||||
*.ipynb linguist-generated
|
||||
13
.github/dependabot.yml
vendored
13
.github/dependabot.yml
vendored
@ -1,13 +0,0 @@
|
||||
# Keep GitHub Actions up to date with GitHub's Dependabot...
|
||||
# https://docs.github.com/en/code-security/dependabot/working-with-dependabot/keeping-your-actions-up-to-date-with-dependabot
|
||||
# https://docs.github.com/en/code-security/dependabot/dependabot-version-updates/configuration-options-for-the-dependabot.yml-file#package-ecosystem
|
||||
version: 2
|
||||
updates:
|
||||
- package-ecosystem: github-actions
|
||||
directory: /
|
||||
groups:
|
||||
github-actions:
|
||||
patterns:
|
||||
- "*" # Group all Actions updates into a single larger pull request
|
||||
schedule:
|
||||
interval: weekly
|
||||
12
.github/workflows/python-publish.yml
vendored
12
.github/workflows/python-publish.yml
vendored
@ -8,23 +8,23 @@ jobs:
|
||||
deploy:
|
||||
runs-on: ubuntu-latest
|
||||
steps:
|
||||
- uses: actions/checkout@v4
|
||||
- uses: actions/checkout@v2
|
||||
- uses: actions-ecosystem/action-regex-match@v2
|
||||
id: regex-match
|
||||
with:
|
||||
text: ${{ github.event.head_commit.message }}
|
||||
regex: '^Release ([^ ]+)'
|
||||
- name: Set up Python
|
||||
uses: actions/setup-python@v5
|
||||
uses: actions/setup-python@v2
|
||||
with:
|
||||
python-version: '3.12'
|
||||
python-version: '3.8'
|
||||
- name: Install dependencies
|
||||
run: |
|
||||
python -m pip install --upgrade pip
|
||||
pip install setuptools wheel twine build
|
||||
pip install setuptools wheel twine
|
||||
- name: Release
|
||||
if: ${{ steps.regex-match.outputs.match != '' }}
|
||||
uses: softprops/action-gh-release@v2
|
||||
uses: softprops/action-gh-release@v1
|
||||
with:
|
||||
tag_name: v${{ steps.regex-match.outputs.group1 }}
|
||||
- name: Build and publish
|
||||
@ -33,5 +33,5 @@ jobs:
|
||||
TWINE_USERNAME: __token__
|
||||
TWINE_PASSWORD: ${{ secrets.PYPI_API_TOKEN }}
|
||||
run: |
|
||||
python -m build --sdist
|
||||
python setup.py sdist
|
||||
twine upload dist/*
|
||||
|
||||
73
.github/workflows/test.yml
vendored
73
.github/workflows/test.yml
vendored
@ -6,74 +6,21 @@ on:
|
||||
pull_request:
|
||||
branches:
|
||||
- main
|
||||
|
||||
jobs:
|
||||
pre-commit:
|
||||
runs-on: ubuntu-latest
|
||||
steps:
|
||||
- uses: actions/checkout@v4
|
||||
- name: Fetch base branch
|
||||
run: git fetch origin ${{ github.base_ref }}
|
||||
- uses: actions/setup-python@v5
|
||||
with:
|
||||
python-version: "3.9"
|
||||
architecture: x64
|
||||
- name: Get pip cache dir
|
||||
id: pip-cache
|
||||
run: |
|
||||
echo "dir=$(pip cache dir)" >> $GITHUB_OUTPUT
|
||||
- name: pip/pre-commit cache
|
||||
uses: actions/cache@v4
|
||||
with:
|
||||
path: |
|
||||
${{ steps.pip-cache.outputs.dir }}
|
||||
~/.cache/pre-commit
|
||||
key: ${{ runner.os }}-pip-pre-commit-${{ hashFiles('**/.pre-commit-config.yaml') }}
|
||||
restore-keys: |
|
||||
${{ runner.os }}-pip-pre-commit
|
||||
- name: pre-commit
|
||||
run: |
|
||||
pip install --upgrade pre-commit
|
||||
pre-commit install --install-hooks
|
||||
pre-commit run --all-files
|
||||
whisper-test:
|
||||
needs: pre-commit
|
||||
runs-on: ubuntu-latest
|
||||
strategy:
|
||||
fail-fast: false
|
||||
matrix:
|
||||
include:
|
||||
- python-version: '3.8'
|
||||
pytorch-version: 1.10.1
|
||||
numpy-requirement: "'numpy<2'"
|
||||
- python-version: '3.8'
|
||||
pytorch-version: 1.13.1
|
||||
numpy-requirement: "'numpy<2'"
|
||||
- python-version: '3.8'
|
||||
pytorch-version: 2.0.1
|
||||
numpy-requirement: "'numpy<2'"
|
||||
- python-version: '3.9'
|
||||
pytorch-version: 2.1.2
|
||||
numpy-requirement: "'numpy<2'"
|
||||
python-version: ['3.8', '3.9', '3.10']
|
||||
pytorch-version: [1.10.2, 1.13.1]
|
||||
exclude:
|
||||
- python-version: '3.10'
|
||||
pytorch-version: 2.2.2
|
||||
numpy-requirement: "'numpy<2'"
|
||||
- python-version: '3.11'
|
||||
pytorch-version: 2.3.1
|
||||
numpy-requirement: "'numpy'"
|
||||
- python-version: '3.12'
|
||||
pytorch-version: 2.4.1
|
||||
numpy-requirement: "'numpy'"
|
||||
- python-version: '3.12'
|
||||
pytorch-version: 2.5.1
|
||||
numpy-requirement: "'numpy'"
|
||||
- python-version: '3.13'
|
||||
pytorch-version: 2.5.1
|
||||
numpy-requirement: "'numpy'"
|
||||
pytorch-version: 1.10.2
|
||||
steps:
|
||||
- uses: conda-incubator/setup-miniconda@v3
|
||||
- run: conda install -n test ffmpeg python=${{ matrix.python-version }}
|
||||
- uses: actions/checkout@v4
|
||||
- uses: conda-incubator/setup-miniconda@v2
|
||||
- run: conda install -n test ffmpeg python=${{ matrix.python-version }} pytorch=${{ matrix.pytorch-version }} cpuonly -c pytorch
|
||||
- uses: actions/checkout@v2
|
||||
- run: echo "$CONDA/envs/test/bin" >> $GITHUB_PATH
|
||||
- run: pip3 install .["dev"] ${{ matrix.numpy-requirement }} torch==${{ matrix.pytorch-version }}+cpu --index-url https://download.pytorch.org/whl/cpu --extra-index-url https://pypi.org/simple
|
||||
- run: pytest --durations=0 -vv -k 'not test_transcribe or test_transcribe[tiny] or test_transcribe[tiny.en]' -m 'not requires_cuda'
|
||||
- run: pip install pytest
|
||||
- run: pip install .
|
||||
- run: pytest -k 'not test_transcribe or test_transcribe[tiny]'
|
||||
|
||||
@ -1,28 +0,0 @@
|
||||
repos:
|
||||
- repo: https://github.com/pre-commit/pre-commit-hooks
|
||||
rev: v5.0.0
|
||||
hooks:
|
||||
- id: check-json
|
||||
- id: end-of-file-fixer
|
||||
types: [file, python]
|
||||
- id: trailing-whitespace
|
||||
types: [file, python]
|
||||
- id: mixed-line-ending
|
||||
- id: check-added-large-files
|
||||
args: [--maxkb=4096]
|
||||
- repo: https://github.com/psf/black
|
||||
rev: 25.1.0
|
||||
hooks:
|
||||
- id: black
|
||||
- repo: https://github.com/pycqa/isort
|
||||
rev: 6.0.0
|
||||
hooks:
|
||||
- id: isort
|
||||
name: isort (python)
|
||||
args: ["--profile", "black", "-l", "88", "--trailing-comma", "--multi-line", "3"]
|
||||
- repo: https://github.com/pycqa/flake8.git
|
||||
rev: 7.1.1
|
||||
hooks:
|
||||
- id: flake8
|
||||
types: [python]
|
||||
args: ["--max-line-length", "88", "--ignore", "E203,E501,W503,W504"]
|
||||
121
CHANGELOG.md
121
CHANGELOG.md
@ -1,121 +0,0 @@
|
||||
# CHANGELOG
|
||||
|
||||
## [v20250625](https://github.com/openai/whisper/releases/tag/v20250625)
|
||||
|
||||
* Fix: Update torch.load to use weights_only=True to prevent security w… ([#2451](https://github.com/openai/whisper/pull/2451))
|
||||
* Fix: Ensure DTW cost tensor is on the same device as input tensor ([#2561](https://github.com/openai/whisper/pull/2561))
|
||||
* docs: updated README to specify translation model limitation ([#2547](https://github.com/openai/whisper/pull/2547))
|
||||
* Fixed triton kernel update to support latest triton versions ([#2588](https://github.com/openai/whisper/pull/2588))
|
||||
* Fix: GitHub display errors for Jupyter notebooks ([#2589](https://github.com/openai/whisper/pull/2589))
|
||||
* Bump the github-actions group with 3 updates ([#2592](https://github.com/openai/whisper/pull/2592))
|
||||
* Keep GitHub Actions up to date with GitHub's Dependabot ([#2486](https://github.com/openai/whisper/pull/2486))
|
||||
* pre-commit: Upgrade black v25.1.0 and isort v6.0.0 ([#2514](https://github.com/openai/whisper/pull/2514))
|
||||
* GitHub Actions: Add Python 3.13 to the testing ([#2487](https://github.com/openai/whisper/pull/2487))
|
||||
* PEP 621: Migrate from setup.py to pyproject.toml ([#2435](https://github.com/openai/whisper/pull/2435))
|
||||
* pre-commit autoupdate && pre-commit run --all-files ([#2484](https://github.com/openai/whisper/pull/2484))
|
||||
* Upgrade GitHub Actions ([#2430](https://github.com/openai/whisper/pull/2430))
|
||||
* Bugfix: Illogical "Avoid computing higher temperatures on no_speech" ([#1903](https://github.com/openai/whisper/pull/1903))
|
||||
* Updating README and doc strings to reflect that n_mels can now be 128 ([#2049](https://github.com/openai/whisper/pull/2049))
|
||||
* fix typo data/README.md ([#2433](https://github.com/openai/whisper/pull/2433))
|
||||
* Update README.md ([#2379](https://github.com/openai/whisper/pull/2379))
|
||||
* Add option to carry initial_prompt with the sliding window ([#2343](https://github.com/openai/whisper/pull/2343))
|
||||
* more pytorch versions in tests ([#2408](https://github.com/openai/whisper/pull/2408))
|
||||
|
||||
## [v20240930](https://github.com/openai/whisper/releases/tag/v20240930)
|
||||
|
||||
* allowing numpy 2 in tests ([#2362](https://github.com/openai/whisper/pull/2362))
|
||||
* large-v3-turbo model ([#2361](https://github.com/openai/whisper/pull/2361))
|
||||
* test on python/pytorch versions up to 3.12 and 2.4.1 ([#2360](https://github.com/openai/whisper/pull/2360))
|
||||
* using sdpa if available ([#2359](https://github.com/openai/whisper/pull/2359))
|
||||
|
||||
## [v20240927](https://github.com/openai/whisper/releases/tag/v20240927)
|
||||
|
||||
* pinning numpy<2 in tests ([#2332](https://github.com/openai/whisper/pull/2332))
|
||||
* Relax triton requirements for compatibility with pytorch 2.4 and newer ([#2307](https://github.com/openai/whisper/pull/2307))
|
||||
* Skip silence around hallucinations ([#1838](https://github.com/openai/whisper/pull/1838))
|
||||
* Fix triton env marker ([#1887](https://github.com/openai/whisper/pull/1887))
|
||||
|
||||
## [v20231117](https://github.com/openai/whisper/releases/tag/v20231117)
|
||||
|
||||
* Relax triton requirements for compatibility with pytorch 2.1 and newer ([#1802](https://github.com/openai/whisper/pull/1802))
|
||||
|
||||
## [v20231106](https://github.com/openai/whisper/releases/tag/v20231106)
|
||||
|
||||
* large-v3 ([#1761](https://github.com/openai/whisper/pull/1761))
|
||||
|
||||
## [v20231105](https://github.com/openai/whisper/releases/tag/v20231105)
|
||||
|
||||
* remove tiktoken pin ([#1759](https://github.com/openai/whisper/pull/1759))
|
||||
* docs: Disambiguation of the term "relative speed" in the README ([#1751](https://github.com/openai/whisper/pull/1751))
|
||||
* allow_pickle=False while loading of mel matrix IN audio.py ([#1511](https://github.com/openai/whisper/pull/1511))
|
||||
* handling transcribe exceptions. ([#1682](https://github.com/openai/whisper/pull/1682))
|
||||
* Add new option to generate subtitles by a specific number of words ([#1729](https://github.com/openai/whisper/pull/1729))
|
||||
* Fix exception when an audio file with no speech is provided ([#1396](https://github.com/openai/whisper/pull/1396))
|
||||
|
||||
## [v20230918](https://github.com/openai/whisper/releases/tag/v20230918)
|
||||
|
||||
* Add .pre-commit-config.yaml ([#1528](https://github.com/openai/whisper/pull/1528))
|
||||
* fix doc of TextDecoder ([#1526](https://github.com/openai/whisper/pull/1526))
|
||||
* Update model-card.md ([#1643](https://github.com/openai/whisper/pull/1643))
|
||||
* word timing tweaks ([#1559](https://github.com/openai/whisper/pull/1559))
|
||||
* Avoid rearranging all caches ([#1483](https://github.com/openai/whisper/pull/1483))
|
||||
* Improve timestamp heuristics. ([#1461](https://github.com/openai/whisper/pull/1461))
|
||||
* fix condition_on_previous_text ([#1224](https://github.com/openai/whisper/pull/1224))
|
||||
* Fix numba depreceation notice ([#1233](https://github.com/openai/whisper/pull/1233))
|
||||
* Updated README.md to provide more insight on BLEU and specific appendices ([#1236](https://github.com/openai/whisper/pull/1236))
|
||||
* Avoid computing higher temperatures on no_speech segments ([#1279](https://github.com/openai/whisper/pull/1279))
|
||||
* Dropped unused execute bit from mel_filters.npz. ([#1254](https://github.com/openai/whisper/pull/1254))
|
||||
* Drop ffmpeg-python dependency and call ffmpeg directly. ([#1242](https://github.com/openai/whisper/pull/1242))
|
||||
* Python 3.11 ([#1171](https://github.com/openai/whisper/pull/1171))
|
||||
* Update decoding.py ([#1219](https://github.com/openai/whisper/pull/1219))
|
||||
* Update decoding.py ([#1155](https://github.com/openai/whisper/pull/1155))
|
||||
* Update README.md to reference tiktoken ([#1105](https://github.com/openai/whisper/pull/1105))
|
||||
* Implement max line width and max line count, and make word highlighting optional ([#1184](https://github.com/openai/whisper/pull/1184))
|
||||
* Squash long words at window and sentence boundaries. ([#1114](https://github.com/openai/whisper/pull/1114))
|
||||
* python-publish.yml: bump actions version to fix node warning ([#1211](https://github.com/openai/whisper/pull/1211))
|
||||
* Update tokenizer.py ([#1163](https://github.com/openai/whisper/pull/1163))
|
||||
|
||||
## [v20230314](https://github.com/openai/whisper/releases/tag/v20230314)
|
||||
|
||||
* abort find_alignment on empty input ([#1090](https://github.com/openai/whisper/pull/1090))
|
||||
* Fix truncated words list when the replacement character is decoded ([#1089](https://github.com/openai/whisper/pull/1089))
|
||||
* fix github language stats getting dominated by jupyter notebook ([#1076](https://github.com/openai/whisper/pull/1076))
|
||||
* Fix alignment between the segments and the list of words ([#1087](https://github.com/openai/whisper/pull/1087))
|
||||
* Use tiktoken ([#1044](https://github.com/openai/whisper/pull/1044))
|
||||
|
||||
## [v20230308](https://github.com/openai/whisper/releases/tag/v20230308)
|
||||
|
||||
* kwargs in decode() for convenience ([#1061](https://github.com/openai/whisper/pull/1061))
|
||||
* fix all_tokens handling that caused more repetitions and discrepancy in JSON ([#1060](https://github.com/openai/whisper/pull/1060))
|
||||
* fix typo in CHANGELOG.md
|
||||
|
||||
## [v20230307](https://github.com/openai/whisper/releases/tag/v20230307)
|
||||
|
||||
* Fix the repetition/hallucination issue identified in #1046 ([#1052](https://github.com/openai/whisper/pull/1052))
|
||||
* Use triton==2.0.0 ([#1053](https://github.com/openai/whisper/pull/1053))
|
||||
* Install triton in x86_64 linux only ([#1051](https://github.com/openai/whisper/pull/1051))
|
||||
* update setup.py to specify python >= 3.8 requirement
|
||||
|
||||
## [v20230306](https://github.com/openai/whisper/releases/tag/v20230306)
|
||||
|
||||
* remove auxiliary audio extension ([#1021](https://github.com/openai/whisper/pull/1021))
|
||||
* apply formatting with `black`, `isort`, and `flake8` ([#1038](https://github.com/openai/whisper/pull/1038))
|
||||
* word-level timestamps in `transcribe()` ([#869](https://github.com/openai/whisper/pull/869))
|
||||
* Decoding improvements ([#1033](https://github.com/openai/whisper/pull/1033))
|
||||
* Update README.md ([#894](https://github.com/openai/whisper/pull/894))
|
||||
* Fix infinite loop caused by incorrect timestamp tokens prediction ([#914](https://github.com/openai/whisper/pull/914))
|
||||
* drop python 3.7 support ([#889](https://github.com/openai/whisper/pull/889))
|
||||
|
||||
## [v20230124](https://github.com/openai/whisper/releases/tag/v20230124)
|
||||
|
||||
* handle printing even if sys.stdout.buffer is not available ([#887](https://github.com/openai/whisper/pull/887))
|
||||
* Add TSV formatted output in transcript, using integer start/end time in milliseconds ([#228](https://github.com/openai/whisper/pull/228))
|
||||
* Added `--output_format` option ([#333](https://github.com/openai/whisper/pull/333))
|
||||
* Handle `XDG_CACHE_HOME` properly for `download_root` ([#864](https://github.com/openai/whisper/pull/864))
|
||||
* use stdout for printing transcription progress ([#867](https://github.com/openai/whisper/pull/867))
|
||||
* Fix bug where mm is mistakenly replaced with hmm in e.g. 20mm ([#659](https://github.com/openai/whisper/pull/659))
|
||||
* print '?' if a letter can't be encoded using the system default encoding ([#859](https://github.com/openai/whisper/pull/859))
|
||||
|
||||
## [v20230117](https://github.com/openai/whisper/releases/tag/v20230117)
|
||||
|
||||
The first versioned release available on [PyPI](https://pypi.org/project/openai-whisper/)
|
||||
@ -2,4 +2,6 @@ include requirements.txt
|
||||
include README.md
|
||||
include LICENSE
|
||||
include whisper/assets/*
|
||||
include whisper/assets/gpt2/*
|
||||
include whisper/assets/multilingual/*
|
||||
include whisper/normalizers/english.json
|
||||
|
||||
63
README.md
63
README.md
@ -5,19 +5,19 @@
|
||||
[[Model card]](https://github.com/openai/whisper/blob/main/model-card.md)
|
||||
[[Colab example]](https://colab.research.google.com/github/openai/whisper/blob/master/notebooks/LibriSpeech.ipynb)
|
||||
|
||||
Whisper is a general-purpose speech recognition model. It is trained on a large dataset of diverse audio and is also a multitasking model that can perform multilingual speech recognition, speech translation, and language identification.
|
||||
Whisper is a general-purpose speech recognition model. It is trained on a large dataset of diverse audio and is also a multi-task model that can perform multilingual speech recognition as well as speech translation and language identification.
|
||||
|
||||
|
||||
## Approach
|
||||
|
||||
|
||||
|
||||
A Transformer sequence-to-sequence model is trained on various speech processing tasks, including multilingual speech recognition, speech translation, spoken language identification, and voice activity detection. These tasks are jointly represented as a sequence of tokens to be predicted by the decoder, allowing a single model to replace many stages of a traditional speech-processing pipeline. The multitask training format uses a set of special tokens that serve as task specifiers or classification targets.
|
||||
A Transformer sequence-to-sequence model is trained on various speech processing tasks, including multilingual speech recognition, speech translation, spoken language identification, and voice activity detection. All of these tasks are jointly represented as a sequence of tokens to be predicted by the decoder, allowing for a single model to replace many different stages of a traditional speech processing pipeline. The multitask training format uses a set of special tokens that serve as task specifiers or classification targets.
|
||||
|
||||
|
||||
## Setup
|
||||
|
||||
We used Python 3.9.9 and [PyTorch](https://pytorch.org/) 1.10.1 to train and test our models, but the codebase is expected to be compatible with Python 3.8-3.11 and recent PyTorch versions. The codebase also depends on a few Python packages, most notably [OpenAI's tiktoken](https://github.com/openai/tiktoken) for their fast tokenizer implementation. You can download and install (or update to) the latest release of Whisper with the following command:
|
||||
We used Python 3.9.9 and [PyTorch](https://pytorch.org/) 1.10.1 to train and test our models, but the codebase is expected to be compatible with Python 3.7 or later and recent PyTorch versions. The codebase also depends on a few Python packages, most notably [HuggingFace Transformers](https://huggingface.co/docs/transformers/index) for their fast tokenizer implementation and [ffmpeg-python](https://github.com/kkroening/ffmpeg-python) for reading audio files. You can download and install (or update to) the latest release of Whisper with the following command:
|
||||
|
||||
pip install -U openai-whisper
|
||||
|
||||
@ -48,7 +48,7 @@ choco install ffmpeg
|
||||
scoop install ffmpeg
|
||||
```
|
||||
|
||||
You may need [`rust`](http://rust-lang.org) installed as well, in case [tiktoken](https://github.com/openai/tiktoken) does not provide a pre-built wheel for your platform. If you see installation errors during the `pip install` command above, please follow the [Getting started page](https://www.rust-lang.org/learn/get-started) to install Rust development environment. Additionally, you may need to configure the `PATH` environment variable, e.g. `export PATH="$HOME/.cargo/bin:$PATH"`. If the installation fails with `No module named 'setuptools_rust'`, you need to install `setuptools_rust`, e.g. by running:
|
||||
You may need [`rust`](http://rust-lang.org) installed as well, in case [tokenizers](https://pypi.org/project/tokenizers/) does not provide a pre-built wheel for your platform. If you see installation errors during the `pip install` command above, please follow the [Getting started page](https://www.rust-lang.org/learn/get-started) to install Rust development environment. Additionally, you may need to configure the `PATH` environment variable, e.g. `export PATH="$HOME/.cargo/bin:$PATH"`. If the installation fails with `No module named 'setuptools_rust'`, you need to install `setuptools_rust`, e.g. by running:
|
||||
|
||||
```bash
|
||||
pip install setuptools-rust
|
||||
@ -57,55 +57,42 @@ pip install setuptools-rust
|
||||
|
||||
## Available models and languages
|
||||
|
||||
There are six model sizes, four with English-only versions, offering speed and accuracy tradeoffs.
|
||||
Below are the names of the available models and their approximate memory requirements and inference speed relative to the large model.
|
||||
The relative speeds below are measured by transcribing English speech on a A100, and the real-world speed may vary significantly depending on many factors including the language, the speaking speed, and the available hardware.
|
||||
There are five model sizes, four with English-only versions, offering speed and accuracy tradeoffs. Below are the names of the available models and their approximate memory requirements and relative speed.
|
||||
|
||||
|
||||
| Size | Parameters | English-only model | Multilingual model | Required VRAM | Relative speed |
|
||||
|:------:|:----------:|:------------------:|:------------------:|:-------------:|:--------------:|
|
||||
| tiny | 39 M | `tiny.en` | `tiny` | ~1 GB | ~10x |
|
||||
| base | 74 M | `base.en` | `base` | ~1 GB | ~7x |
|
||||
| small | 244 M | `small.en` | `small` | ~2 GB | ~4x |
|
||||
| tiny | 39 M | `tiny.en` | `tiny` | ~1 GB | ~32x |
|
||||
| base | 74 M | `base.en` | `base` | ~1 GB | ~16x |
|
||||
| small | 244 M | `small.en` | `small` | ~2 GB | ~6x |
|
||||
| medium | 769 M | `medium.en` | `medium` | ~5 GB | ~2x |
|
||||
| large | 1550 M | N/A | `large` | ~10 GB | 1x |
|
||||
| turbo | 809 M | N/A | `turbo` | ~6 GB | ~8x |
|
||||
|
||||
The `.en` models for English-only applications tend to perform better, especially for the `tiny.en` and `base.en` models. We observed that the difference becomes less significant for the `small.en` and `medium.en` models.
|
||||
Additionally, the `turbo` model is an optimized version of `large-v3` that offers faster transcription speed with a minimal degradation in accuracy.
|
||||
For English-only applications, the `.en` models tend to perform better, especially for the `tiny.en` and `base.en` models. We observed that the difference becomes less significant for the `small.en` and `medium.en` models.
|
||||
|
||||
Whisper's performance varies widely depending on the language. The figure below shows a WER (Word Error Rate) breakdown by languages of Fleurs dataset, using the `large-v2` model. More WER and BLEU scores corresponding to the other models and datasets can be found in Appendix D in [the paper](https://arxiv.org/abs/2212.04356). The smaller is better.
|
||||
|
||||
|
||||
|
||||
Whisper's performance varies widely depending on the language. The figure below shows a performance breakdown of `large-v3` and `large-v2` models by language, using WERs (word error rates) or CER (character error rates, shown in *Italic*) evaluated on the Common Voice 15 and Fleurs datasets. Additional WER/CER metrics corresponding to the other models and datasets can be found in Appendix D.1, D.2, and D.4 of [the paper](https://arxiv.org/abs/2212.04356), as well as the BLEU (Bilingual Evaluation Understudy) scores for translation in Appendix D.3.
|
||||
|
||||
|
||||
|
||||
## Command-line usage
|
||||
|
||||
The following command will transcribe speech in audio files, using the `turbo` model:
|
||||
The following command will transcribe speech in audio files, using the `medium` model:
|
||||
|
||||
```bash
|
||||
whisper audio.flac audio.mp3 audio.wav --model turbo
|
||||
```
|
||||
whisper audio.flac audio.mp3 audio.wav --model medium
|
||||
|
||||
The default setting (which selects the `turbo` model) works well for transcribing English. However, **the `turbo` model is not trained for translation tasks**. If you need to **translate non-English speech into English**, use one of the **multilingual models** (`tiny`, `base`, `small`, `medium`, `large`) instead of `turbo`.
|
||||
The default setting (which selects the `small` model) works well for transcribing English. To transcribe an audio file containing non-English speech, you can specify the language using the `--language` option:
|
||||
|
||||
For example, to transcribe an audio file containing non-English speech, you can specify the language:
|
||||
whisper japanese.wav --language Japanese
|
||||
|
||||
```bash
|
||||
whisper japanese.wav --language Japanese
|
||||
```
|
||||
Adding `--task translate` will translate the speech into English:
|
||||
|
||||
To **translate** speech into English, use:
|
||||
|
||||
```bash
|
||||
whisper japanese.wav --model medium --language Japanese --task translate
|
||||
```
|
||||
|
||||
> **Note:** The `turbo` model will return the original language even if `--task translate` is specified. Use `medium` or `large` for the best translation results.
|
||||
whisper japanese.wav --language Japanese --task translate
|
||||
|
||||
Run the following to view all available options:
|
||||
|
||||
```bash
|
||||
whisper --help
|
||||
```
|
||||
whisper --help
|
||||
|
||||
See [tokenizer.py](https://github.com/openai/whisper/blob/main/whisper/tokenizer.py) for the list of all available languages.
|
||||
|
||||
@ -117,7 +104,7 @@ Transcription can also be performed within Python:
|
||||
```python
|
||||
import whisper
|
||||
|
||||
model = whisper.load_model("turbo")
|
||||
model = whisper.load_model("base")
|
||||
result = model.transcribe("audio.mp3")
|
||||
print(result["text"])
|
||||
```
|
||||
@ -129,14 +116,14 @@ Below is an example usage of `whisper.detect_language()` and `whisper.decode()`
|
||||
```python
|
||||
import whisper
|
||||
|
||||
model = whisper.load_model("turbo")
|
||||
model = whisper.load_model("base")
|
||||
|
||||
# load audio and pad/trim it to fit 30 seconds
|
||||
audio = whisper.load_audio("audio.mp3")
|
||||
audio = whisper.pad_or_trim(audio)
|
||||
|
||||
# make log-Mel spectrogram and move to the same device as the model
|
||||
mel = whisper.log_mel_spectrogram(audio, n_mels=model.dims.n_mels).to(model.device)
|
||||
mel = whisper.log_mel_spectrogram(audio).to(model.device)
|
||||
|
||||
# detect the spoken language
|
||||
_, probs = model.detect_language(mel)
|
||||
@ -157,4 +144,4 @@ Please use the [🙌 Show and tell](https://github.com/openai/whisper/discussion
|
||||
|
||||
## License
|
||||
|
||||
Whisper's code and model weights are released under the MIT License. See [LICENSE](https://github.com/openai/whisper/blob/main/LICENSE) for further details.
|
||||
The code and the model weights of Whisper are released under the MIT License. See [LICENSE](https://github.com/openai/whisper/blob/main/LICENSE) for further details.
|
||||
|
||||
@ -45,7 +45,7 @@ We downloaded the [CHiME-5 dataset](https://spandh.dcs.shef.ac.uk//chime_challen
|
||||
|
||||
### AMI-IHM, AMI-SDM1
|
||||
|
||||
We preprocessed the [AMI Corpus](https://groups.inf.ed.ac.uk/ami/corpus/overview.shtml) by following the stage 0 and 2 of the [s5b recipe](https://github.com/kaldi-asr/kaldi/tree/master/egs/ami/s5b).
|
||||
We preprocessed the [AMI Corpus](https://groups.inf.ed.ac.uk/ami/corpus/overview.shtml) by following the stage 0 ad 2 of the [s5b recipe](https://github.com/kaldi-asr/kaldi/tree/master/egs/ami/s5b).
|
||||
|
||||
|
||||
## Long-form English-only datasets
|
||||
|
||||
File diff suppressed because it is too large
Load Diff
|
Before Width: | Height: | Size: 272 KiB After Width: | Height: | Size: 100 KiB |
@ -16,15 +16,13 @@ The Whisper models are trained for speech recognition and translation tasks, cap
|
||||
| small | 244 M | ✓ | ✓ |
|
||||
| medium | 769 M | ✓ | ✓ |
|
||||
| large | 1550 M | | ✓ |
|
||||
| turbo | 798 M | | ✓ |
|
||||
|
||||
In December 2022, we [released an improved large model named `large-v2`](https://github.com/openai/whisper/discussions/661), and `large-v3` in November 2023.
|
||||
Additionally, we've added a `turbo` model in September 2024 which is optimized for inference speed.
|
||||
In December 2022, we [released an improved large model named `large-v2`](https://github.com/openai/whisper/discussions/661).
|
||||
|
||||
|
||||
### Release date
|
||||
|
||||
September 2022 (original series), December 2022 (`large-v2`), November 2023 (`large-v3`), September 2024 (`large-v3-turbo`)
|
||||
September 2022 (original series) and December 2022 (`large-v2`)
|
||||
|
||||
### Model type
|
||||
|
||||
@ -39,7 +37,7 @@ Sequence-to-sequence ASR (automatic speech recognition) and speech translation m
|
||||
|
||||
### Evaluated Use
|
||||
|
||||
The primary intended users of these models are AI researchers studying the robustness, generalization, capabilities, biases, and constraints of the current model. However, Whisper is also potentially quite useful as an ASR solution for developers, especially for English speech recognition. We recognize that once models are released, it is impossible to restrict access to only “intended” uses or to draw reasonable guidelines around what is or is not research.
|
||||
The primary intended users of these models are AI researchers studying robustness, generalization, capabilities, biases, and constraints of the current model. However, Whisper is also potentially quite useful as an ASR solution for developers, especially for English speech recognition. We recognize that once models are released, it is impossible to restrict access to only “intended” uses or to draw reasonable guidelines around what is or is not research.
|
||||
|
||||
The models are primarily trained and evaluated on ASR and speech translation to English tasks. They show strong ASR results in ~10 languages. They may exhibit additional capabilities, particularly if fine-tuned on certain tasks like voice activity detection, speaker classification, or speaker diarization but have not been robustly evaluated in these areas. We strongly recommend that users perform robust evaluations of the models in a particular context and domain before deploying them.
|
||||
|
||||
@ -55,17 +53,17 @@ As discussed in [the accompanying paper](https://arxiv.org/abs/2212.04356), we s
|
||||
|
||||
## Performance and Limitations
|
||||
|
||||
Our studies show that, over many existing ASR systems, the models exhibit improved robustness to accents, background noise, and technical language, as well as zero-shot translation from multiple languages into English; and that accuracy on speech recognition and translation is near the state-of-the-art level.
|
||||
Our studies show that, over many existing ASR systems, the models exhibit improved robustness to accents, background noise, technical language, as well as zero shot translation from multiple languages into English; and that accuracy on speech recognition and translation is near the state-of-the-art level.
|
||||
|
||||
However, because the models are trained in a weakly supervised manner using large-scale noisy data, the predictions may include texts that are not actually spoken in the audio input (i.e. hallucination). We hypothesize that this happens because, given their general knowledge of language, the models combine trying to predict the next word in audio with trying to transcribe the audio itself.
|
||||
|
||||
Our models perform unevenly across languages, and we observe lower accuracy on low-resource and/or low-discoverability languages or languages where we have less training data. The models also exhibit disparate performance on different accents and dialects of particular languages, which may include a higher word error rate across speakers of different genders, races, ages, or other demographic criteria. Our full evaluation results are presented in [the paper accompanying this release](https://arxiv.org/abs/2212.04356).
|
||||
Our models perform unevenly across languages, and we observe lower accuracy on low-resource and/or low-discoverability languages or languages where we have less training data. The models also exhibit disparate performance on different accents and dialects of particular languages, which may include higher word error rate across speakers of different genders, races, ages, or other demographic criteria. Our full evaluation results are presented in [the paper accompanying this release](https://arxiv.org/abs/2212.04356).
|
||||
|
||||
In addition, the sequence-to-sequence architecture of the model makes it prone to generating repetitive texts, which can be mitigated to some degree by beam search and temperature scheduling but not perfectly. Further analysis of these limitations is provided in [the paper](https://arxiv.org/abs/2212.04356). It is likely that this behavior and hallucinations may be worse in lower-resource and/or lower-discoverability languages.
|
||||
In addition, the sequence-to-sequence architecture of the model makes it prone to generating repetitive texts, which can be mitigated to some degree by beam search and temperature scheduling but not perfectly. Further analysis on these limitations are provided in [the paper](https://arxiv.org/abs/2212.04356). It is likely that this behavior and hallucinations may be worse on lower-resource and/or lower-discoverability languages.
|
||||
|
||||
|
||||
## Broader Implications
|
||||
|
||||
We anticipate that Whisper models’ transcription capabilities may be used for improving accessibility tools. While Whisper models cannot be used for real-time transcription out of the box – their speed and size suggest that others may be able to build applications on top of them that allow for near-real-time speech recognition and translation. The real value of beneficial applications built on top of Whisper models suggests that the disparate performance of these models may have real economic implications.
|
||||
|
||||
There are also potential dual-use concerns that come with releasing Whisper. While we hope the technology will be used primarily for beneficial purposes, making ASR technology more accessible could enable more actors to build capable surveillance technologies or scale up existing surveillance efforts, as the speed and accuracy allow for affordable automatic transcription and translation of large volumes of audio communication. Moreover, these models may have some capabilities to recognize specific individuals out of the box, which in turn presents safety concerns related both to dual use and disparate performance. In practice, we expect that the cost of transcription is not the limiting factor of scaling up surveillance projects.
|
||||
There are also potential dual use concerns that come with releasing Whisper. While we hope the technology will be used primarily for beneficial purposes, making ASR technology more accessible could enable more actors to build capable surveillance technologies or scale up existing surveillance efforts, as the speed and accuracy allow for affordable automatic transcription and translation of large volumes of audio communication. Moreover, these models may have some capabilities to recognize specific individuals out of the box, which in turn presents safety concerns related both to dual use and disparate performance. In practice, we expect that the cost of transcription is not the limiting factor of scaling up surveillance projects.
|
||||
|
||||
@ -949,8 +949,7 @@
|
||||
"style": "IPY_MODEL_039b53f2702c4179af7e0548018d0588",
|
||||
"value": " 164/164 [05:08<00:00, 1.86s/it]"
|
||||
}
|
||||
},
|
||||
"state": {}
|
||||
}
|
||||
}
|
||||
}
|
||||
},
|
||||
|
||||
@ -4219,8 +4219,7 @@
|
||||
"_view_name": "StyleView",
|
||||
"description_width": ""
|
||||
}
|
||||
},
|
||||
"state": {}
|
||||
}
|
||||
}
|
||||
}
|
||||
},
|
||||
|
||||
@ -1,54 +0,0 @@
|
||||
[build-system]
|
||||
build-backend = "setuptools.build_meta"
|
||||
|
||||
requires = [ "setuptools>=61.2" ]
|
||||
|
||||
[project]
|
||||
name = "openai-whisper"
|
||||
description = "Robust Speech Recognition via Large-Scale Weak Supervision"
|
||||
readme.content-type = "text/markdown"
|
||||
readme.file = "README.md"
|
||||
license = { text = "MIT" }
|
||||
authors = [ { name = "OpenAI" } ]
|
||||
requires-python = ">=3.8"
|
||||
classifiers = [
|
||||
"Programming Language :: Python :: 3 :: Only",
|
||||
"Programming Language :: Python :: 3.8",
|
||||
"Programming Language :: Python :: 3.9",
|
||||
"Programming Language :: Python :: 3.10",
|
||||
"Programming Language :: Python :: 3.11",
|
||||
"Programming Language :: Python :: 3.12",
|
||||
"Programming Language :: Python :: 3.13",
|
||||
]
|
||||
dynamic = [ "version" ]
|
||||
dependencies = [
|
||||
"more-itertools",
|
||||
"numba",
|
||||
"numpy",
|
||||
"tiktoken",
|
||||
"torch",
|
||||
"tqdm",
|
||||
"triton>=2; (platform_machine=='x86_64' and sys_platform=='linux') or sys_platform=='linux2'",
|
||||
]
|
||||
optional-dependencies.dev = [ "black", "flake8", "isort", "pytest", "scipy" ]
|
||||
urls = { Homepage = "https://github.com/openai/whisper" }
|
||||
scripts.whisper = "whisper.transcribe:cli"
|
||||
|
||||
[tool.setuptools]
|
||||
py-modules = [ "whisper" ]
|
||||
include-package-data = true
|
||||
|
||||
[tool.setuptools.dynamic]
|
||||
version = { attr = "whisper.version.__version__" }
|
||||
|
||||
[tool.setuptools.packages.find]
|
||||
exclude = [ "tests*" ]
|
||||
namespaces = false
|
||||
|
||||
[tool.black]
|
||||
|
||||
[tool.isort]
|
||||
profile = "black"
|
||||
include_trailing_comma = true
|
||||
line_length = 88
|
||||
multi_line_output = 3
|
||||
@ -1,7 +1,6 @@
|
||||
numba
|
||||
numpy
|
||||
torch
|
||||
tqdm
|
||||
more-itertools
|
||||
tiktoken
|
||||
triton>=2.0.0;platform_machine=="x86_64" and sys_platform=="linux" or sys_platform=="linux2"
|
||||
transformers>=4.19.0
|
||||
ffmpeg-python==0.2.0
|
||||
|
||||
36
setup.py
Normal file
36
setup.py
Normal file
@ -0,0 +1,36 @@
|
||||
import os
|
||||
|
||||
import pkg_resources
|
||||
from setuptools import setup, find_packages
|
||||
|
||||
|
||||
def read_version(fname="whisper/version.py"):
|
||||
exec(compile(open(fname, encoding="utf-8").read(), fname, "exec"))
|
||||
return locals()["__version__"]
|
||||
|
||||
|
||||
setup(
|
||||
name="openai-whisper",
|
||||
py_modules=["whisper"],
|
||||
version=read_version(),
|
||||
description="Robust Speech Recognition via Large-Scale Weak Supervision",
|
||||
long_description=open("README.md", encoding="utf-8").read(),
|
||||
long_description_content_type="text/markdown",
|
||||
readme="README.md",
|
||||
python_requires=">=3.7",
|
||||
author="OpenAI",
|
||||
url="https://github.com/openai/whisper",
|
||||
license="MIT",
|
||||
packages=find_packages(exclude=["tests*"]),
|
||||
install_requires=[
|
||||
str(r)
|
||||
for r in pkg_resources.parse_requirements(
|
||||
open(os.path.join(os.path.dirname(__file__), "requirements.txt"))
|
||||
)
|
||||
],
|
||||
entry_points={
|
||||
"console_scripts": ["whisper=whisper.transcribe:cli"],
|
||||
},
|
||||
include_package_data=True,
|
||||
extras_require={"dev": ["pytest"]},
|
||||
)
|
||||
@ -1,14 +0,0 @@
|
||||
import random as rand
|
||||
|
||||
import numpy
|
||||
import pytest
|
||||
|
||||
|
||||
def pytest_configure(config):
|
||||
config.addinivalue_line("markers", "requires_cuda")
|
||||
|
||||
|
||||
@pytest.fixture
|
||||
def random():
|
||||
rand.seed(42)
|
||||
numpy.random.seed(42)
|
||||
@ -2,7 +2,7 @@ import os.path
|
||||
|
||||
import numpy as np
|
||||
|
||||
from whisper.audio import SAMPLE_RATE, load_audio, log_mel_spectrogram
|
||||
from whisper.audio import load_audio, log_mel_spectrogram, SAMPLE_RATE
|
||||
|
||||
|
||||
def test_audio():
|
||||
|
||||
@ -1,10 +1,7 @@
|
||||
import pytest
|
||||
|
||||
from whisper.normalizers import EnglishTextNormalizer
|
||||
from whisper.normalizers.english import (
|
||||
EnglishNumberNormalizer,
|
||||
EnglishSpellingNormalizer,
|
||||
)
|
||||
from whisper.normalizers.english import EnglishNumberNormalizer, EnglishSpellingNormalizer
|
||||
|
||||
|
||||
@pytest.mark.parametrize("std", [EnglishNumberNormalizer(), EnglishTextNormalizer()])
|
||||
@ -87,7 +84,6 @@ def test_text_normalizer():
|
||||
assert std("he's like") == "he is like"
|
||||
assert std("she's been like") == "she has been like"
|
||||
assert std("10km") == "10 km"
|
||||
assert std("10mm") == "10 mm"
|
||||
assert std("RC232") == "rc 232"
|
||||
|
||||
assert (
|
||||
|
||||
@ -1,96 +0,0 @@
|
||||
import numpy as np
|
||||
import pytest
|
||||
import scipy.ndimage
|
||||
import torch
|
||||
|
||||
from whisper.timing import dtw_cpu, dtw_cuda, median_filter
|
||||
|
||||
sizes = [
|
||||
(10, 20),
|
||||
(32, 16),
|
||||
(123, 1500),
|
||||
(234, 189),
|
||||
]
|
||||
shapes = [
|
||||
(10,),
|
||||
(1, 15),
|
||||
(4, 5, 345),
|
||||
(6, 12, 240, 512),
|
||||
]
|
||||
|
||||
|
||||
@pytest.mark.parametrize("N, M", sizes)
|
||||
def test_dtw(N: int, M: int):
|
||||
steps = np.concatenate([np.zeros(N - 1), np.ones(M - 1)])
|
||||
np.random.shuffle(steps)
|
||||
x = np.random.random((N, M)).astype(np.float32)
|
||||
|
||||
i, j, k = 0, 0, 0
|
||||
trace = []
|
||||
while True:
|
||||
x[i, j] -= 1
|
||||
trace.append((i, j))
|
||||
|
||||
if k == len(steps):
|
||||
break
|
||||
|
||||
if k + 1 < len(steps) and steps[k] != steps[k + 1]:
|
||||
i += 1
|
||||
j += 1
|
||||
k += 2
|
||||
continue
|
||||
|
||||
if steps[k] == 0:
|
||||
i += 1
|
||||
if steps[k] == 1:
|
||||
j += 1
|
||||
k += 1
|
||||
|
||||
trace = np.array(trace).T
|
||||
dtw_trace = dtw_cpu(x)
|
||||
|
||||
assert np.allclose(trace, dtw_trace)
|
||||
|
||||
|
||||
@pytest.mark.requires_cuda
|
||||
@pytest.mark.parametrize("N, M", sizes)
|
||||
def test_dtw_cuda_equivalence(N: int, M: int):
|
||||
x_numpy = np.random.randn(N, M).astype(np.float32)
|
||||
x_cuda = torch.from_numpy(x_numpy).cuda()
|
||||
|
||||
trace_cpu = dtw_cpu(x_numpy)
|
||||
trace_cuda = dtw_cuda(x_cuda)
|
||||
|
||||
assert np.allclose(trace_cpu, trace_cuda)
|
||||
|
||||
|
||||
@pytest.mark.parametrize("shape", shapes)
|
||||
def test_median_filter(shape):
|
||||
x = torch.randn(*shape)
|
||||
|
||||
for filter_width in [3, 5, 7, 13]:
|
||||
filtered = median_filter(x, filter_width)
|
||||
|
||||
# using np.pad to reflect-pad, because Scipy's behavior is different near the edges.
|
||||
pad_width = filter_width // 2
|
||||
padded_x = np.pad(
|
||||
x, [(0, 0)] * (x.ndim - 1) + [(pad_width, pad_width)], mode="reflect"
|
||||
)
|
||||
scipy_filtered = scipy.ndimage.median_filter(
|
||||
padded_x, [1] * (x.ndim - 1) + [filter_width]
|
||||
)
|
||||
scipy_filtered = scipy_filtered[..., pad_width:-pad_width]
|
||||
|
||||
assert np.allclose(filtered, scipy_filtered)
|
||||
|
||||
|
||||
@pytest.mark.requires_cuda
|
||||
@pytest.mark.parametrize("shape", shapes)
|
||||
def test_median_filter_equivalence(shape):
|
||||
x = torch.randn(*shape)
|
||||
|
||||
for filter_width in [3, 5, 7, 13]:
|
||||
filtered_cpu = median_filter(x, filter_width)
|
||||
filtered_gpu = median_filter(x.cuda(), filter_width).cpu()
|
||||
|
||||
assert np.allclose(filtered_cpu, filtered_gpu)
|
||||
@ -1,17 +1,7 @@
|
||||
import pytest
|
||||
|
||||
from whisper.tokenizer import get_tokenizer
|
||||
|
||||
|
||||
@pytest.mark.parametrize("multilingual", [True, False])
|
||||
def test_tokenizer(multilingual):
|
||||
tokenizer = get_tokenizer(multilingual=False)
|
||||
assert tokenizer.sot in tokenizer.sot_sequence
|
||||
assert len(tokenizer.all_language_codes) == len(tokenizer.all_language_tokens)
|
||||
assert all(c < tokenizer.timestamp_begin for c in tokenizer.all_language_tokens)
|
||||
|
||||
|
||||
def test_multilingual_tokenizer():
|
||||
def test_tokenizer():
|
||||
gpt2_tokenizer = get_tokenizer(multilingual=False)
|
||||
multilingual_tokenizer = get_tokenizer(multilingual=True)
|
||||
|
||||
@ -22,13 +12,3 @@ def test_multilingual_tokenizer():
|
||||
assert gpt2_tokenizer.decode(gpt2_tokens) == text
|
||||
assert multilingual_tokenizer.decode(multilingual_tokens) == text
|
||||
assert len(gpt2_tokens) > len(multilingual_tokens)
|
||||
|
||||
|
||||
def test_split_on_unicode():
|
||||
multilingual_tokenizer = get_tokenizer(multilingual=True)
|
||||
|
||||
tokens = [8404, 871, 287, 6, 246, 526, 3210, 20378]
|
||||
words, word_tokens = multilingual_tokenizer.split_tokens_on_unicode(tokens)
|
||||
|
||||
assert words == [" elle", " est", " l", "'", "\ufffd", "é", "rit", "oire"]
|
||||
assert word_tokens == [[8404], [871], [287], [6], [246], [526], [3210], [20378]]
|
||||
|
||||
@ -4,7 +4,6 @@ import pytest
|
||||
import torch
|
||||
|
||||
import whisper
|
||||
from whisper.tokenizer import get_tokenizer
|
||||
|
||||
|
||||
@pytest.mark.parametrize("model_name", whisper.available_models())
|
||||
@ -14,29 +13,10 @@ def test_transcribe(model_name: str):
|
||||
audio_path = os.path.join(os.path.dirname(__file__), "jfk.flac")
|
||||
|
||||
language = "en" if model_name.endswith(".en") else None
|
||||
result = model.transcribe(
|
||||
audio_path, language=language, temperature=0.0, word_timestamps=True
|
||||
)
|
||||
result = model.transcribe(audio_path, language=language, temperature=0.0)
|
||||
assert result["language"] == "en"
|
||||
assert result["text"] == "".join([s["text"] for s in result["segments"]])
|
||||
|
||||
transcription = result["text"].lower()
|
||||
assert "my fellow americans" in transcription
|
||||
assert "your country" in transcription
|
||||
assert "do for you" in transcription
|
||||
|
||||
tokenizer = get_tokenizer(model.is_multilingual, num_languages=model.num_languages)
|
||||
all_tokens = [t for s in result["segments"] for t in s["tokens"]]
|
||||
assert tokenizer.decode(all_tokens) == result["text"]
|
||||
assert tokenizer.decode_with_timestamps(all_tokens).startswith("<|0.00|>")
|
||||
|
||||
timing_checked = False
|
||||
for segment in result["segments"]:
|
||||
for timing in segment["words"]:
|
||||
assert timing["start"] < timing["end"]
|
||||
if timing["word"].strip(" ,") == "Americans":
|
||||
assert timing["start"] <= 1.8
|
||||
assert timing["end"] >= 1.8
|
||||
timing_checked = True
|
||||
|
||||
assert timing_checked
|
||||
|
||||
@ -10,10 +10,11 @@ from tqdm import tqdm
|
||||
|
||||
from .audio import load_audio, log_mel_spectrogram, pad_or_trim
|
||||
from .decoding import DecodingOptions, DecodingResult, decode, detect_language
|
||||
from .model import ModelDimensions, Whisper
|
||||
from .model import Whisper, ModelDimensions
|
||||
from .transcribe import transcribe
|
||||
from .version import __version__
|
||||
|
||||
|
||||
_MODELS = {
|
||||
"tiny.en": "https://openaipublic.azureedge.net/main/whisper/models/d3dd57d32accea0b295c96e26691aa14d8822fac7d9d27d5dc00b4ca2826dd03/tiny.en.pt",
|
||||
"tiny": "https://openaipublic.azureedge.net/main/whisper/models/65147644a518d12f04e32d6f3b26facc3f8dd46e5390956a9424a650c0ce22b9/tiny.pt",
|
||||
@ -25,29 +26,7 @@ _MODELS = {
|
||||
"medium": "https://openaipublic.azureedge.net/main/whisper/models/345ae4da62f9b3d59415adc60127b97c714f32e89e936602e85993674d08dcb1/medium.pt",
|
||||
"large-v1": "https://openaipublic.azureedge.net/main/whisper/models/e4b87e7e0bf463eb8e6956e646f1e277e901512310def2c24bf0e11bd3c28e9a/large-v1.pt",
|
||||
"large-v2": "https://openaipublic.azureedge.net/main/whisper/models/81f7c96c852ee8fc832187b0132e569d6c3065a3252ed18e56effd0b6a73e524/large-v2.pt",
|
||||
"large-v3": "https://openaipublic.azureedge.net/main/whisper/models/e5b1a55b89c1367dacf97e3e19bfd829a01529dbfdeefa8caeb59b3f1b81dadb/large-v3.pt",
|
||||
"large": "https://openaipublic.azureedge.net/main/whisper/models/e5b1a55b89c1367dacf97e3e19bfd829a01529dbfdeefa8caeb59b3f1b81dadb/large-v3.pt",
|
||||
"large-v3-turbo": "https://openaipublic.azureedge.net/main/whisper/models/aff26ae408abcba5fbf8813c21e62b0941638c5f6eebfb145be0c9839262a19a/large-v3-turbo.pt",
|
||||
"turbo": "https://openaipublic.azureedge.net/main/whisper/models/aff26ae408abcba5fbf8813c21e62b0941638c5f6eebfb145be0c9839262a19a/large-v3-turbo.pt",
|
||||
}
|
||||
|
||||
# base85-encoded (n_layers, n_heads) boolean arrays indicating the cross-attention heads that are
|
||||
# highly correlated to the word-level timing, i.e. the alignment between audio and text tokens.
|
||||
_ALIGNMENT_HEADS = {
|
||||
"tiny.en": b"ABzY8J1N>@0{>%R00Bk>$p{7v037`oCl~+#00",
|
||||
"tiny": b"ABzY8bu8Lr0{>%RKn9Fp%m@SkK7Kt=7ytkO",
|
||||
"base.en": b"ABzY8;40c<0{>%RzzG;p*o+Vo09|#PsxSZm00",
|
||||
"base": b"ABzY8KQ!870{>%RzyTQH3`Q^yNP!>##QT-<FaQ7m",
|
||||
"small.en": b"ABzY8>?_)10{>%RpeA61k&I|OI3I$65C{;;pbCHh0B{qLQ;+}v00",
|
||||
"small": b"ABzY8DmU6=0{>%Rpa?J`kvJ6qF(V^F86#Xh7JUGMK}P<N0000",
|
||||
"medium.en": b"ABzY8usPae0{>%R7<zz_OvQ{)4kMa0BMw6u5rT}kRKX;$NfYBv00*Hl@qhsU00",
|
||||
"medium": b"ABzY8B0Jh+0{>%R7}kK1fFL7w6%<-Pf*t^=N)Qr&0RR9",
|
||||
"large-v1": b"ABzY8r9j$a0{>%R7#4sLmoOs{s)o3~84-RPdcFk!JR<kSfC2yj",
|
||||
"large-v2": b"ABzY8zd+h!0{>%R7=D0pU<_bnWW*tkYAhobTNnu$jnkEkXqp)j;w1Tzk)UH3X%SZd&fFZ2fC2yj",
|
||||
"large-v3": b"ABzY8gWO1E0{>%R7(9S+Kn!D~%ngiGaR?*L!iJG9p-nab0JQ=-{D1-g00",
|
||||
"large": b"ABzY8gWO1E0{>%R7(9S+Kn!D~%ngiGaR?*L!iJG9p-nab0JQ=-{D1-g00",
|
||||
"large-v3-turbo": b"ABzY8j^C+e0{>%RARaKHP%t(lGR*)0g!tONPyhe`",
|
||||
"turbo": b"ABzY8j^C+e0{>%RARaKHP%t(lGR*)0g!tONPyhe`",
|
||||
"large": "https://openaipublic.azureedge.net/main/whisper/models/81f7c96c852ee8fc832187b0132e569d6c3065a3252ed18e56effd0b6a73e524/large-v2.pt",
|
||||
}
|
||||
|
||||
|
||||
@ -66,18 +45,10 @@ def _download(url: str, root: str, in_memory: bool) -> Union[bytes, str]:
|
||||
if hashlib.sha256(model_bytes).hexdigest() == expected_sha256:
|
||||
return model_bytes if in_memory else download_target
|
||||
else:
|
||||
warnings.warn(
|
||||
f"{download_target} exists, but the SHA256 checksum does not match; re-downloading the file"
|
||||
)
|
||||
warnings.warn(f"{download_target} exists, but the SHA256 checksum does not match; re-downloading the file")
|
||||
|
||||
with urllib.request.urlopen(url) as source, open(download_target, "wb") as output:
|
||||
with tqdm(
|
||||
total=int(source.info().get("Content-Length")),
|
||||
ncols=80,
|
||||
unit="iB",
|
||||
unit_scale=True,
|
||||
unit_divisor=1024,
|
||||
) as loop:
|
||||
with tqdm(total=int(source.info().get("Content-Length")), ncols=80, unit='iB', unit_scale=True, unit_divisor=1024) as loop:
|
||||
while True:
|
||||
buffer = source.read(8192)
|
||||
if not buffer:
|
||||
@ -88,9 +59,7 @@ def _download(url: str, root: str, in_memory: bool) -> Union[bytes, str]:
|
||||
|
||||
model_bytes = open(download_target, "rb").read()
|
||||
if hashlib.sha256(model_bytes).hexdigest() != expected_sha256:
|
||||
raise RuntimeError(
|
||||
"Model has been downloaded but the SHA256 checksum does not not match. Please retry loading the model."
|
||||
)
|
||||
raise RuntimeError("Model has been downloaded but the SHA256 checksum does not not match. Please retry loading the model.")
|
||||
|
||||
return model_bytes if in_memory else download_target
|
||||
|
||||
@ -100,12 +69,7 @@ def available_models() -> List[str]:
|
||||
return list(_MODELS.keys())
|
||||
|
||||
|
||||
def load_model(
|
||||
name: str,
|
||||
device: Optional[Union[str, torch.device]] = None,
|
||||
download_root: str = None,
|
||||
in_memory: bool = False,
|
||||
) -> Whisper:
|
||||
def load_model(name: str, device: Optional[Union[str, torch.device]] = None, download_root: str = None, in_memory: bool = False) -> Whisper:
|
||||
"""
|
||||
Load a Whisper ASR model
|
||||
|
||||
@ -130,32 +94,24 @@ def load_model(
|
||||
if device is None:
|
||||
device = "cuda" if torch.cuda.is_available() else "cpu"
|
||||
if download_root is None:
|
||||
default = os.path.join(os.path.expanduser("~"), ".cache")
|
||||
download_root = os.path.join(os.getenv("XDG_CACHE_HOME", default), "whisper")
|
||||
download_root = os.getenv(
|
||||
"XDG_CACHE_HOME",
|
||||
os.path.join(os.path.expanduser("~"), ".cache", "whisper")
|
||||
)
|
||||
|
||||
if name in _MODELS:
|
||||
checkpoint_file = _download(_MODELS[name], download_root, in_memory)
|
||||
alignment_heads = _ALIGNMENT_HEADS[name]
|
||||
elif os.path.isfile(name):
|
||||
checkpoint_file = open(name, "rb").read() if in_memory else name
|
||||
alignment_heads = None
|
||||
else:
|
||||
raise RuntimeError(
|
||||
f"Model {name} not found; available models = {available_models()}"
|
||||
)
|
||||
raise RuntimeError(f"Model {name} not found; available models = {available_models()}")
|
||||
|
||||
with (
|
||||
io.BytesIO(checkpoint_file) if in_memory else open(checkpoint_file, "rb")
|
||||
) as fp:
|
||||
kwargs = {"weights_only": True} if torch.__version__ >= "1.13" else {}
|
||||
checkpoint = torch.load(fp, map_location=device, **kwargs)
|
||||
with (io.BytesIO(checkpoint_file) if in_memory else open(checkpoint_file, "rb")) as fp:
|
||||
checkpoint = torch.load(fp, map_location=device)
|
||||
del checkpoint_file
|
||||
|
||||
dims = ModelDimensions(**checkpoint["dims"])
|
||||
model = Whisper(dims)
|
||||
model.load_state_dict(checkpoint["model_state_dict"])
|
||||
|
||||
if alignment_heads is not None:
|
||||
model.set_alignment_heads(alignment_heads)
|
||||
|
||||
return model.to(device)
|
||||
|
||||
@ -1,3 +1,4 @@
|
||||
from .transcribe import cli
|
||||
|
||||
|
||||
cli()
|
||||
|
||||
File diff suppressed because it is too large
Load Diff
50001
whisper/assets/gpt2/merges.txt
Normal file
50001
whisper/assets/gpt2/merges.txt
Normal file
File diff suppressed because it is too large
Load Diff
1
whisper/assets/gpt2/special_tokens_map.json
Normal file
1
whisper/assets/gpt2/special_tokens_map.json
Normal file
@ -0,0 +1 @@
|
||||
{"bos_token": "<|endoftext|>", "eos_token": "<|endoftext|>", "unk_token": "<|endoftext|>"}
|
||||
1
whisper/assets/gpt2/tokenizer_config.json
Normal file
1
whisper/assets/gpt2/tokenizer_config.json
Normal file
@ -0,0 +1 @@
|
||||
{"unk_token": "<|endoftext|>", "bos_token": "<|endoftext|>", "eos_token": "<|endoftext|>", "add_prefix_space": false, "model_max_length": 1024, "special_tokens_map_file": null, "name_or_path": "gpt2", "tokenizer_class": "GPT2Tokenizer"}
|
||||
1
whisper/assets/gpt2/vocab.json
Normal file
1
whisper/assets/gpt2/vocab.json
Normal file
File diff suppressed because one or more lines are too long
BIN
whisper/assets/mel_filters.npz
Normal file → Executable file
BIN
whisper/assets/mel_filters.npz
Normal file → Executable file
Binary file not shown.
File diff suppressed because it is too large
Load Diff
1
whisper/assets/multilingual/added_tokens.json
Normal file
1
whisper/assets/multilingual/added_tokens.json
Normal file
@ -0,0 +1 @@
|
||||
{"<|endoftext|>": 50257}
|
||||
50000
whisper/assets/multilingual/merges.txt
Normal file
50000
whisper/assets/multilingual/merges.txt
Normal file
File diff suppressed because it is too large
Load Diff
1
whisper/assets/multilingual/special_tokens_map.json
Normal file
1
whisper/assets/multilingual/special_tokens_map.json
Normal file
@ -0,0 +1 @@
|
||||
{"bos_token": "<|endoftext|>", "eos_token": "<|endoftext|>", "unk_token": "<|endoftext|>"}
|
||||
1
whisper/assets/multilingual/tokenizer_config.json
Normal file
1
whisper/assets/multilingual/tokenizer_config.json
Normal file
@ -0,0 +1 @@
|
||||
{"unk_token": {"content": "<|endoftext|>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true, "__type": "AddedToken"}, "bos_token": {"content": "<|endoftext|>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true, "__type": "AddedToken"}, "eos_token": {"content": "<|endoftext|>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true, "__type": "AddedToken"}, "add_prefix_space": false, "model_max_length": 1024, "special_tokens_map_file": null, "name_or_path": "multilingual", "errors": "replace", "tokenizer_class": "GPT2Tokenizer"}
|
||||
1
whisper/assets/multilingual/vocab.json
Normal file
1
whisper/assets/multilingual/vocab.json
Normal file
File diff suppressed because one or more lines are too long
@ -1,8 +1,8 @@
|
||||
import os
|
||||
from functools import lru_cache
|
||||
from subprocess import CalledProcessError, run
|
||||
from typing import Optional, Union
|
||||
from typing import Union
|
||||
|
||||
import ffmpeg
|
||||
import numpy as np
|
||||
import torch
|
||||
import torch.nn.functional as F
|
||||
@ -12,14 +12,11 @@ from .utils import exact_div
|
||||
# hard-coded audio hyperparameters
|
||||
SAMPLE_RATE = 16000
|
||||
N_FFT = 400
|
||||
N_MELS = 80
|
||||
HOP_LENGTH = 160
|
||||
CHUNK_LENGTH = 30
|
||||
N_SAMPLES = CHUNK_LENGTH * SAMPLE_RATE # 480000 samples in a 30-second chunk
|
||||
N_FRAMES = exact_div(N_SAMPLES, HOP_LENGTH) # 3000 frames in a mel spectrogram input
|
||||
|
||||
N_SAMPLES_PER_TOKEN = HOP_LENGTH * 2 # the initial convolutions has stride 2
|
||||
FRAMES_PER_SECOND = exact_div(SAMPLE_RATE, HOP_LENGTH) # 10ms per audio frame
|
||||
TOKENS_PER_SECOND = exact_div(SAMPLE_RATE, N_SAMPLES_PER_TOKEN) # 20ms per audio token
|
||||
N_SAMPLES = CHUNK_LENGTH * SAMPLE_RATE # 480000: number of samples in a chunk
|
||||
N_FRAMES = exact_div(N_SAMPLES, HOP_LENGTH) # 3000: number of frames in a mel spectrogram input
|
||||
|
||||
|
||||
def load_audio(file: str, sr: int = SAMPLE_RATE):
|
||||
@ -38,25 +35,15 @@ def load_audio(file: str, sr: int = SAMPLE_RATE):
|
||||
-------
|
||||
A NumPy array containing the audio waveform, in float32 dtype.
|
||||
"""
|
||||
|
||||
# This launches a subprocess to decode audio while down-mixing
|
||||
# and resampling as necessary. Requires the ffmpeg CLI in PATH.
|
||||
# fmt: off
|
||||
cmd = [
|
||||
"ffmpeg",
|
||||
"-nostdin",
|
||||
"-threads", "0",
|
||||
"-i", file,
|
||||
"-f", "s16le",
|
||||
"-ac", "1",
|
||||
"-acodec", "pcm_s16le",
|
||||
"-ar", str(sr),
|
||||
"-"
|
||||
]
|
||||
# fmt: on
|
||||
try:
|
||||
out = run(cmd, capture_output=True, check=True).stdout
|
||||
except CalledProcessError as e:
|
||||
# This launches a subprocess to decode audio while down-mixing and resampling as necessary.
|
||||
# Requires the ffmpeg CLI and `ffmpeg-python` package to be installed.
|
||||
out, _ = (
|
||||
ffmpeg.input(file, threads=0)
|
||||
.output("-", format="s16le", acodec="pcm_s16le", ac=1, ar=sr)
|
||||
.run(cmd=["ffmpeg", "-nostdin"], capture_stdout=True, capture_stderr=True)
|
||||
)
|
||||
except ffmpeg.Error as e:
|
||||
raise RuntimeError(f"Failed to load audio: {e.stderr.decode()}") from e
|
||||
|
||||
return np.frombuffer(out, np.int16).flatten().astype(np.float32) / 32768.0
|
||||
@ -68,9 +55,7 @@ def pad_or_trim(array, length: int = N_SAMPLES, *, axis: int = -1):
|
||||
"""
|
||||
if torch.is_tensor(array):
|
||||
if array.shape[axis] > length:
|
||||
array = array.index_select(
|
||||
dim=axis, index=torch.arange(length, device=array.device)
|
||||
)
|
||||
array = array.index_select(dim=axis, index=torch.arange(length, device=array.device))
|
||||
|
||||
if array.shape[axis] < length:
|
||||
pad_widths = [(0, 0)] * array.ndim
|
||||
@ -89,7 +74,7 @@ def pad_or_trim(array, length: int = N_SAMPLES, *, axis: int = -1):
|
||||
|
||||
|
||||
@lru_cache(maxsize=None)
|
||||
def mel_filters(device, n_mels: int) -> torch.Tensor:
|
||||
def mel_filters(device, n_mels: int = N_MELS) -> torch.Tensor:
|
||||
"""
|
||||
load the mel filterbank matrix for projecting STFT into a Mel spectrogram.
|
||||
Allows decoupling librosa dependency; saved using:
|
||||
@ -97,22 +82,14 @@ def mel_filters(device, n_mels: int) -> torch.Tensor:
|
||||
np.savez_compressed(
|
||||
"mel_filters.npz",
|
||||
mel_80=librosa.filters.mel(sr=16000, n_fft=400, n_mels=80),
|
||||
mel_128=librosa.filters.mel(sr=16000, n_fft=400, n_mels=128),
|
||||
)
|
||||
"""
|
||||
assert n_mels in {80, 128}, f"Unsupported n_mels: {n_mels}"
|
||||
|
||||
filters_path = os.path.join(os.path.dirname(__file__), "assets", "mel_filters.npz")
|
||||
with np.load(filters_path, allow_pickle=False) as f:
|
||||
assert n_mels == 80, f"Unsupported n_mels: {n_mels}"
|
||||
with np.load(os.path.join(os.path.dirname(__file__), "assets", "mel_filters.npz")) as f:
|
||||
return torch.from_numpy(f[f"mel_{n_mels}"]).to(device)
|
||||
|
||||
|
||||
def log_mel_spectrogram(
|
||||
audio: Union[str, np.ndarray, torch.Tensor],
|
||||
n_mels: int = 80,
|
||||
padding: int = 0,
|
||||
device: Optional[Union[str, torch.device]] = None,
|
||||
):
|
||||
def log_mel_spectrogram(audio: Union[str, np.ndarray, torch.Tensor], n_mels: int = N_MELS):
|
||||
"""
|
||||
Compute the log-Mel spectrogram of
|
||||
|
||||
@ -122,17 +99,11 @@ def log_mel_spectrogram(
|
||||
The path to audio or either a NumPy array or Tensor containing the audio waveform in 16 kHz
|
||||
|
||||
n_mels: int
|
||||
The number of Mel-frequency filters, only 80 and 128 are supported
|
||||
|
||||
padding: int
|
||||
Number of zero samples to pad to the right
|
||||
|
||||
device: Optional[Union[str, torch.device]]
|
||||
If given, the audio tensor is moved to this device before STFT
|
||||
The number of Mel-frequency filters, only 80 is supported
|
||||
|
||||
Returns
|
||||
-------
|
||||
torch.Tensor, shape = (n_mels, n_frames)
|
||||
torch.Tensor, shape = (80, n_frames)
|
||||
A Tensor that contains the Mel spectrogram
|
||||
"""
|
||||
if not torch.is_tensor(audio):
|
||||
@ -140,10 +111,6 @@ def log_mel_spectrogram(
|
||||
audio = load_audio(audio)
|
||||
audio = torch.from_numpy(audio)
|
||||
|
||||
if device is not None:
|
||||
audio = audio.to(device)
|
||||
if padding > 0:
|
||||
audio = F.pad(audio, (0, padding))
|
||||
window = torch.hann_window(N_FFT).to(audio.device)
|
||||
stft = torch.stft(audio, N_FFT, HOP_LENGTH, window=window, return_complex=True)
|
||||
magnitudes = stft[..., :-1].abs() ** 2
|
||||
|
||||
@ -1,5 +1,5 @@
|
||||
from dataclasses import dataclass, field, replace
|
||||
from typing import TYPE_CHECKING, Dict, Iterable, List, Optional, Sequence, Tuple, Union
|
||||
from dataclasses import dataclass, field
|
||||
from typing import Dict, List, Tuple, Iterable, Optional, Sequence, Union, TYPE_CHECKING
|
||||
|
||||
import numpy as np
|
||||
import torch
|
||||
@ -16,9 +16,7 @@ if TYPE_CHECKING:
|
||||
|
||||
|
||||
@torch.no_grad()
|
||||
def detect_language(
|
||||
model: "Whisper", mel: Tensor, tokenizer: Tokenizer = None
|
||||
) -> Tuple[Tensor, List[dict]]:
|
||||
def detect_language(model: "Whisper", mel: Tensor, tokenizer: Tokenizer = None) -> Tuple[Tensor, List[dict]]:
|
||||
"""
|
||||
Detect the spoken language in the audio, and return them as list of strings, along with the ids
|
||||
of the most probable language tokens and the probability distribution over all language tokens.
|
||||
@ -32,16 +30,9 @@ def detect_language(
|
||||
list of dictionaries containing the probability distribution over all languages.
|
||||
"""
|
||||
if tokenizer is None:
|
||||
tokenizer = get_tokenizer(
|
||||
model.is_multilingual, num_languages=model.num_languages
|
||||
)
|
||||
if (
|
||||
tokenizer.language is None
|
||||
or tokenizer.language_token not in tokenizer.sot_sequence
|
||||
):
|
||||
raise ValueError(
|
||||
"This model doesn't have language tokens so it can't perform lang id"
|
||||
)
|
||||
tokenizer = get_tokenizer(model.is_multilingual)
|
||||
if tokenizer.language is None or tokenizer.language_token not in tokenizer.sot_sequence:
|
||||
raise ValueError(f"This model doesn't have language tokens so it can't perform lang id")
|
||||
|
||||
single = mel.ndim == 2
|
||||
if single:
|
||||
@ -79,36 +70,31 @@ def detect_language(
|
||||
|
||||
@dataclass(frozen=True)
|
||||
class DecodingOptions:
|
||||
# whether to perform X->X "transcribe" or X->English "translate"
|
||||
task: str = "transcribe"
|
||||
|
||||
# language that the audio is in; uses detected language if None
|
||||
language: Optional[str] = None
|
||||
task: str = "transcribe" # whether to perform X->X "transcribe" or X->English "translate"
|
||||
language: Optional[str] = None # language that the audio is in; uses detected language if None
|
||||
|
||||
# sampling-related options
|
||||
temperature: float = 0.0
|
||||
sample_len: Optional[int] = None # maximum number of tokens to sample
|
||||
best_of: Optional[int] = None # number of independent sample trajectories, if t > 0
|
||||
beam_size: Optional[int] = None # number of beams in beam search, if t == 0
|
||||
patience: Optional[float] = None # patience in beam search (arxiv:2204.05424)
|
||||
best_of: Optional[int] = None # number of independent samples to collect, when t > 0
|
||||
beam_size: Optional[int] = None # number of beams in beam search, when t == 0
|
||||
patience: Optional[float] = None # patience in beam search (https://arxiv.org/abs/2204.05424)
|
||||
|
||||
# "alpha" in Google NMT, or None for length norm, when ranking generations
|
||||
# to select which to return among the beams or best-of-N samples
|
||||
length_penalty: Optional[float] = None
|
||||
# options for ranking generations (either beams or best-of-N samples)
|
||||
length_penalty: Optional[float] = None # "alpha" in Google NMT, None defaults to length norm
|
||||
|
||||
# text or tokens to feed as the prompt or the prefix; for more info:
|
||||
# https://github.com/openai/whisper/discussions/117#discussioncomment-3727051
|
||||
prompt: Optional[Union[str, List[int]]] = None # for the previous context
|
||||
prefix: Optional[Union[str, List[int]]] = None # to prefix the current context
|
||||
# prompt, prefix, and token suppression
|
||||
prompt: Optional[Union[str, List[int]]] = None # text or tokens for the previous context
|
||||
prefix: Optional[Union[str, List[int]]] = None # text or tokens to prefix the current context
|
||||
suppress_blank: bool = True # this will suppress blank outputs
|
||||
|
||||
# list of tokens ids (or comma-separated token ids) to suppress
|
||||
# "-1" will suppress a set of symbols as defined in `tokenizer.non_speech_tokens()`
|
||||
suppress_tokens: Optional[Union[str, Iterable[int]]] = "-1"
|
||||
suppress_blank: bool = True # this will suppress blank outputs
|
||||
|
||||
# timestamp sampling options
|
||||
without_timestamps: bool = False # use <|notimestamps|> to sample text tokens only
|
||||
max_initial_timestamp: Optional[float] = 1.0
|
||||
without_timestamps: bool = False # use <|notimestamps|> to sample text tokens only
|
||||
max_initial_timestamp: Optional[float] = 1.0 # the initial timestamp cannot be later than this
|
||||
|
||||
# implementation details
|
||||
fp16: bool = True # use fp16 for most of the calculation
|
||||
@ -148,10 +134,6 @@ class PyTorchInference(Inference):
|
||||
self.kv_cache = {}
|
||||
self.hooks = []
|
||||
|
||||
key_modules = [block.attn.key for block in self.model.decoder.blocks]
|
||||
value_modules = [block.attn.value for block in self.model.decoder.blocks]
|
||||
self.kv_modules = key_modules + value_modules
|
||||
|
||||
def logits(self, tokens: Tensor, audio_features: Tensor) -> Tensor:
|
||||
if not self.kv_cache:
|
||||
self.kv_cache, self.hooks = self.model.install_kv_cache_hooks()
|
||||
@ -170,16 +152,13 @@ class PyTorchInference(Inference):
|
||||
self.hooks = []
|
||||
|
||||
def rearrange_kv_cache(self, source_indices):
|
||||
if source_indices != list(range(len(source_indices))):
|
||||
for module in self.kv_modules:
|
||||
# update the key/value cache to contain the selected sequences
|
||||
self.kv_cache[module] = self.kv_cache[module][source_indices].detach()
|
||||
for module, tensor in self.kv_cache.items():
|
||||
# update the key/value cache to contain the selected sequences
|
||||
self.kv_cache[module] = tensor[source_indices].detach()
|
||||
|
||||
|
||||
class SequenceRanker:
|
||||
def rank(
|
||||
self, tokens: List[List[Tensor]], sum_logprobs: List[List[float]]
|
||||
) -> List[int]:
|
||||
def rank(self, tokens: List[List[Tensor]], sum_logprobs: List[List[float]]) -> List[int]:
|
||||
"""
|
||||
Given a list of groups of samples and their cumulative log probabilities,
|
||||
return the indices of the samples in each group to select as the final result
|
||||
@ -217,9 +196,7 @@ class TokenDecoder:
|
||||
def reset(self):
|
||||
"""Initialize any stateful variables for decoding a new sequence"""
|
||||
|
||||
def update(
|
||||
self, tokens: Tensor, logits: Tensor, sum_logprobs: Tensor
|
||||
) -> Tuple[Tensor, bool]:
|
||||
def update(self, tokens: Tensor, logits: Tensor, sum_logprobs: Tensor) -> Tuple[Tensor, bool]:
|
||||
"""Specify how to select the next token, based on the current trace and logits
|
||||
|
||||
Parameters
|
||||
@ -274,13 +251,12 @@ class GreedyDecoder(TokenDecoder):
|
||||
self.temperature = temperature
|
||||
self.eot = eot
|
||||
|
||||
def update(
|
||||
self, tokens: Tensor, logits: Tensor, sum_logprobs: Tensor
|
||||
) -> Tuple[Tensor, bool]:
|
||||
if self.temperature == 0:
|
||||
def update(self, tokens: Tensor, logits: Tensor, sum_logprobs: Tensor) -> Tuple[Tensor, bool]:
|
||||
temperature = self.temperature
|
||||
if temperature == 0:
|
||||
next_tokens = logits.argmax(dim=-1)
|
||||
else:
|
||||
next_tokens = Categorical(logits=logits / self.temperature).sample()
|
||||
next_tokens = Categorical(logits=logits / temperature).sample()
|
||||
|
||||
logprobs = F.log_softmax(logits.float(), dim=-1)
|
||||
current_logprobs = logprobs[torch.arange(logprobs.shape[0]), next_tokens]
|
||||
@ -299,13 +275,7 @@ class GreedyDecoder(TokenDecoder):
|
||||
|
||||
|
||||
class BeamSearchDecoder(TokenDecoder):
|
||||
def __init__(
|
||||
self,
|
||||
beam_size: int,
|
||||
eot: int,
|
||||
inference: Inference,
|
||||
patience: Optional[float] = None,
|
||||
):
|
||||
def __init__(self, beam_size: int, eot: int, inference: Inference, patience: Optional[float] = None):
|
||||
self.beam_size = beam_size
|
||||
self.eot = eot
|
||||
self.inference = inference
|
||||
@ -313,16 +283,12 @@ class BeamSearchDecoder(TokenDecoder):
|
||||
self.max_candidates: int = round(beam_size * self.patience)
|
||||
self.finished_sequences = None
|
||||
|
||||
assert (
|
||||
self.max_candidates > 0
|
||||
), f"Invalid beam size ({beam_size}) or patience ({patience})"
|
||||
assert self.max_candidates > 0, f"Invalid beam size ({beam_size}) or patience ({patience})"
|
||||
|
||||
def reset(self):
|
||||
self.finished_sequences = None
|
||||
|
||||
def update(
|
||||
self, tokens: Tensor, logits: Tensor, sum_logprobs: Tensor
|
||||
) -> Tuple[Tensor, bool]:
|
||||
def update(self, tokens: Tensor, logits: Tensor, sum_logprobs: Tensor) -> Tuple[Tensor, bool]:
|
||||
if tokens.shape[0] % self.beam_size != 0:
|
||||
raise ValueError(f"{tokens.shape}[0] % {self.beam_size} != 0")
|
||||
|
||||
@ -366,9 +332,7 @@ class BeamSearchDecoder(TokenDecoder):
|
||||
|
||||
# add newly finished sequences to self.finished_sequences
|
||||
assert len(self.finished_sequences) == len(finished_sequences)
|
||||
for previously_finished, newly_finished in zip(
|
||||
self.finished_sequences, finished_sequences
|
||||
):
|
||||
for previously_finished, newly_finished in zip(self.finished_sequences, finished_sequences):
|
||||
for seq in sorted(newly_finished, key=newly_finished.get, reverse=True):
|
||||
if len(previously_finished) >= self.max_candidates:
|
||||
break # the candidate list is full
|
||||
@ -376,8 +340,7 @@ class BeamSearchDecoder(TokenDecoder):
|
||||
|
||||
# mark as completed if all audio has enough number of samples
|
||||
completed = all(
|
||||
len(sequences) >= self.max_candidates
|
||||
for sequences in self.finished_sequences
|
||||
len(sequences) >= self.max_candidates for sequences in self.finished_sequences
|
||||
)
|
||||
return tokens, completed
|
||||
|
||||
@ -385,9 +348,7 @@ class BeamSearchDecoder(TokenDecoder):
|
||||
# collect all finished sequences, including patience, and add unfinished ones if not enough
|
||||
sum_logprobs = sum_logprobs.cpu()
|
||||
for i, sequences in enumerate(self.finished_sequences):
|
||||
if (
|
||||
len(sequences) < self.beam_size
|
||||
): # when not enough sequences are finished
|
||||
if len(sequences) < self.beam_size: # when not enough sequences are finished
|
||||
for j in list(np.argsort(sum_logprobs[i]))[::-1]:
|
||||
sequence = preceding_tokens[i, j].tolist() + [self.eot]
|
||||
sequences[tuple(sequence)] = sum_logprobs[i][j].item()
|
||||
@ -395,8 +356,7 @@ class BeamSearchDecoder(TokenDecoder):
|
||||
break
|
||||
|
||||
tokens: List[List[Tensor]] = [
|
||||
[torch.tensor(seq) for seq in sequences.keys()]
|
||||
for sequences in self.finished_sequences
|
||||
[torch.tensor(seq) for seq in sequences.keys()] for sequences in self.finished_sequences
|
||||
]
|
||||
sum_logprobs: List[List[float]] = [
|
||||
list(sequences.values()) for sequences in self.finished_sequences
|
||||
@ -440,10 +400,7 @@ class SuppressTokens(LogitFilter):
|
||||
|
||||
class ApplyTimestampRules(LogitFilter):
|
||||
def __init__(
|
||||
self,
|
||||
tokenizer: Tokenizer,
|
||||
sample_begin: int,
|
||||
max_initial_timestamp_index: Optional[int],
|
||||
self, tokenizer: Tokenizer, sample_begin: int, max_initial_timestamp_index: Optional[int]
|
||||
):
|
||||
self.tokenizer = tokenizer
|
||||
self.sample_begin = sample_begin
|
||||
@ -456,14 +413,9 @@ class ApplyTimestampRules(LogitFilter):
|
||||
|
||||
# timestamps have to appear in pairs, except directly before EOT; mask logits accordingly
|
||||
for k in range(tokens.shape[0]):
|
||||
sampled_tokens = tokens[k, self.sample_begin :]
|
||||
seq = [t for t in sampled_tokens.tolist()]
|
||||
last_was_timestamp = (
|
||||
len(seq) >= 1 and seq[-1] >= self.tokenizer.timestamp_begin
|
||||
)
|
||||
penultimate_was_timestamp = (
|
||||
len(seq) < 2 or seq[-2] >= self.tokenizer.timestamp_begin
|
||||
)
|
||||
seq = [t for t in tokens[k, self.sample_begin :].tolist()]
|
||||
last_was_timestamp = len(seq) >= 1 and seq[-1] >= self.tokenizer.timestamp_begin
|
||||
penultimate_was_timestamp = len(seq) < 2 or seq[-2] >= self.tokenizer.timestamp_begin
|
||||
|
||||
if last_was_timestamp:
|
||||
if penultimate_was_timestamp: # has to be non-timestamp
|
||||
@ -471,35 +423,19 @@ class ApplyTimestampRules(LogitFilter):
|
||||
else: # cannot be normal text tokens
|
||||
logits[k, : self.tokenizer.eot] = -np.inf
|
||||
|
||||
timestamps = sampled_tokens[
|
||||
sampled_tokens.ge(self.tokenizer.timestamp_begin)
|
||||
]
|
||||
if timestamps.numel() > 0:
|
||||
# timestamps shouldn't decrease; forbid timestamp tokens smaller than the last
|
||||
# also force each segment to have a nonzero length, to prevent infinite looping
|
||||
if last_was_timestamp and not penultimate_was_timestamp:
|
||||
timestamp_last = timestamps[-1]
|
||||
else:
|
||||
timestamp_last = timestamps[-1] + 1
|
||||
logits[k, self.tokenizer.timestamp_begin : timestamp_last] = -np.inf
|
||||
|
||||
if tokens.shape[1] == self.sample_begin:
|
||||
# suppress generating non-timestamp tokens at the beginning
|
||||
logits[:, : self.tokenizer.timestamp_begin] = -np.inf
|
||||
|
||||
# apply the `max_initial_timestamp` option
|
||||
if self.max_initial_timestamp_index is not None:
|
||||
last_allowed = (
|
||||
self.tokenizer.timestamp_begin + self.max_initial_timestamp_index
|
||||
)
|
||||
last_allowed = self.tokenizer.timestamp_begin + self.max_initial_timestamp_index
|
||||
logits[:, last_allowed + 1 :] = -np.inf
|
||||
|
||||
# if sum of probability over timestamps is above any other token, sample timestamp
|
||||
logprobs = F.log_softmax(logits.float(), dim=-1)
|
||||
for k in range(tokens.shape[0]):
|
||||
timestamp_logprob = logprobs[k, self.tokenizer.timestamp_begin :].logsumexp(
|
||||
dim=-1
|
||||
)
|
||||
timestamp_logprob = logprobs[k, self.tokenizer.timestamp_begin :].logsumexp(dim=-1)
|
||||
max_text_token_logprob = logprobs[k, : self.tokenizer.timestamp_begin].max()
|
||||
if timestamp_logprob > max_text_token_logprob:
|
||||
logits[k, : self.tokenizer.timestamp_begin] = -np.inf
|
||||
@ -515,12 +451,7 @@ class DecodingTask:
|
||||
self.model = model
|
||||
|
||||
language = options.language or "en"
|
||||
tokenizer = get_tokenizer(
|
||||
model.is_multilingual,
|
||||
num_languages=model.num_languages,
|
||||
language=language,
|
||||
task=options.task,
|
||||
)
|
||||
tokenizer = get_tokenizer(model.is_multilingual, language=language, task=options.task)
|
||||
self.tokenizer: Tokenizer = tokenizer
|
||||
self.options: DecodingOptions = self._verify_options(options)
|
||||
|
||||
@ -560,13 +491,9 @@ class DecodingTask:
|
||||
precision = CHUNK_LENGTH / model.dims.n_audio_ctx # usually 0.02 seconds
|
||||
max_initial_timestamp_index = None
|
||||
if options.max_initial_timestamp:
|
||||
max_initial_timestamp_index = round(
|
||||
self.options.max_initial_timestamp / precision
|
||||
)
|
||||
max_initial_timestamp_index = round(self.options.max_initial_timestamp / precision)
|
||||
self.logit_filters.append(
|
||||
ApplyTimestampRules(
|
||||
tokenizer, self.sample_begin, max_initial_timestamp_index
|
||||
)
|
||||
ApplyTimestampRules(tokenizer, self.sample_begin, max_initial_timestamp_index)
|
||||
)
|
||||
|
||||
def _verify_options(self, options: DecodingOptions) -> DecodingOptions:
|
||||
@ -577,38 +504,30 @@ class DecodingTask:
|
||||
raise ValueError("best_of with greedy sampling (T=0) is not compatible")
|
||||
if options.patience is not None and options.beam_size is None:
|
||||
raise ValueError("patience requires beam_size to be given")
|
||||
if options.length_penalty is not None and not (
|
||||
0 <= options.length_penalty <= 1
|
||||
):
|
||||
if options.length_penalty is not None and not (0 <= options.length_penalty <= 1):
|
||||
raise ValueError("length_penalty (alpha) should be a value between 0 and 1")
|
||||
|
||||
return options
|
||||
|
||||
def _get_initial_tokens(self) -> Tuple[int]:
|
||||
tokens = list(self.sot_sequence)
|
||||
prefix = self.options.prefix
|
||||
prompt = self.options.prompt
|
||||
|
||||
if prefix := self.options.prefix:
|
||||
if prefix:
|
||||
prefix_tokens = (
|
||||
self.tokenizer.encode(" " + prefix.strip())
|
||||
if isinstance(prefix, str)
|
||||
else prefix
|
||||
self.tokenizer.encode(" " + prefix.strip()) if isinstance(prefix, str) else prefix
|
||||
)
|
||||
if self.sample_len is not None:
|
||||
max_prefix_len = self.n_ctx // 2 - self.sample_len
|
||||
prefix_tokens = prefix_tokens[-max_prefix_len:]
|
||||
tokens = tokens + prefix_tokens
|
||||
|
||||
if prompt := self.options.prompt:
|
||||
if prompt:
|
||||
prompt_tokens = (
|
||||
self.tokenizer.encode(" " + prompt.strip())
|
||||
if isinstance(prompt, str)
|
||||
else prompt
|
||||
)
|
||||
tokens = (
|
||||
[self.tokenizer.sot_prev]
|
||||
+ prompt_tokens[-(self.n_ctx // 2 - 1) :]
|
||||
+ tokens
|
||||
self.tokenizer.encode(" " + prompt.strip()) if isinstance(prompt, str) else prompt
|
||||
)
|
||||
tokens = [self.tokenizer.sot_prev] + prompt_tokens[-(self.n_ctx // 2 - 1) :] + tokens
|
||||
|
||||
return tuple(tokens)
|
||||
|
||||
@ -627,13 +546,7 @@ class DecodingTask:
|
||||
assert isinstance(suppress_tokens, list), "suppress_tokens must be a list"
|
||||
|
||||
suppress_tokens.extend(
|
||||
[
|
||||
self.tokenizer.transcribe,
|
||||
self.tokenizer.translate,
|
||||
self.tokenizer.sot,
|
||||
self.tokenizer.sot_prev,
|
||||
self.tokenizer.sot_lm,
|
||||
]
|
||||
[self.tokenizer.sot, self.tokenizer.sot_prev, self.tokenizer.sot_lm]
|
||||
)
|
||||
if self.tokenizer.no_speech is not None:
|
||||
# no-speech probability is collected separately
|
||||
@ -645,21 +558,14 @@ class DecodingTask:
|
||||
if self.options.fp16:
|
||||
mel = mel.half()
|
||||
|
||||
if mel.shape[-2:] == (
|
||||
self.model.dims.n_audio_ctx,
|
||||
self.model.dims.n_audio_state,
|
||||
):
|
||||
if mel.shape[-2:] == (self.model.dims.n_audio_ctx, self.model.dims.n_audio_state):
|
||||
# encoded audio features are given; skip audio encoding
|
||||
audio_features = mel
|
||||
else:
|
||||
audio_features = self.model.encoder(mel)
|
||||
|
||||
if audio_features.dtype != (
|
||||
torch.float16 if self.options.fp16 else torch.float32
|
||||
):
|
||||
return TypeError(
|
||||
f"audio_features has an incorrect dtype: {audio_features.dtype}"
|
||||
)
|
||||
if audio_features.dtype != (torch.float16 if self.options.fp16 else torch.float32):
|
||||
return TypeError(f"audio_features has an incorrect dtype: {audio_features.dtype}")
|
||||
|
||||
return audio_features
|
||||
|
||||
@ -668,9 +574,7 @@ class DecodingTask:
|
||||
lang_probs = None
|
||||
|
||||
if self.options.language is None or self.options.task == "lang_id":
|
||||
lang_tokens, lang_probs = self.model.detect_language(
|
||||
audio_features, self.tokenizer
|
||||
)
|
||||
lang_tokens, lang_probs = self.model.detect_language(audio_features, self.tokenizer)
|
||||
languages = [max(probs, key=probs.get) for probs in lang_probs]
|
||||
if self.options.language is None:
|
||||
tokens[:, self.sot_index + 1] = lang_tokens # write language tokens
|
||||
@ -678,6 +582,7 @@ class DecodingTask:
|
||||
return languages, lang_probs
|
||||
|
||||
def _main_loop(self, audio_features: Tensor, tokens: Tensor):
|
||||
assert audio_features.shape[0] == tokens.shape[0]
|
||||
n_batch = tokens.shape[0]
|
||||
sum_logprobs: Tensor = torch.zeros(n_batch, device=audio_features.device)
|
||||
no_speech_probs = [np.nan] * n_batch
|
||||
@ -686,9 +591,7 @@ class DecodingTask:
|
||||
for i in range(self.sample_len):
|
||||
logits = self.inference.logits(tokens, audio_features)
|
||||
|
||||
if (
|
||||
i == 0 and self.tokenizer.no_speech is not None
|
||||
): # save no_speech_probs
|
||||
if i == 0 and self.tokenizer.no_speech is not None: # save no_speech_probs
|
||||
probs_at_sot = logits[:, self.sot_index].float().softmax(dim=-1)
|
||||
no_speech_probs = probs_at_sot[:, self.tokenizer.no_speech].tolist()
|
||||
|
||||
@ -722,15 +625,12 @@ class DecodingTask:
|
||||
languages, language_probs = self._detect_language(audio_features, tokens)
|
||||
if self.options.task == "lang_id":
|
||||
return [
|
||||
DecodingResult(
|
||||
audio_features=features, language=language, language_probs=probs
|
||||
)
|
||||
for features, language, probs in zip(
|
||||
audio_features, languages, language_probs
|
||||
)
|
||||
DecodingResult(audio_features=features, language=language, language_probs=probs)
|
||||
for features, language, probs in zip(audio_features, languages, language_probs)
|
||||
]
|
||||
|
||||
# repeat text tensors by the group size, for beam search or best-of-n sampling
|
||||
# repeat the audio & text tensors by the group size, for beam search or best-of-n sampling
|
||||
audio_features = audio_features.repeat_interleave(self.n_group, dim=0)
|
||||
tokens = tokens.repeat_interleave(self.n_group, dim=0).to(audio_features.device)
|
||||
|
||||
# call the main sampling loop
|
||||
@ -747,8 +647,7 @@ class DecodingTask:
|
||||
# get the final candidates for each group, and slice between the first sampled token and EOT
|
||||
tokens, sum_logprobs = self.decoder.finalize(tokens, sum_logprobs)
|
||||
tokens: List[List[Tensor]] = [
|
||||
[t[self.sample_begin : (t == tokenizer.eot).nonzero()[0, 0]] for t in s]
|
||||
for s in tokens
|
||||
[t[self.sample_begin : (t == tokenizer.eot).nonzero()[0, 0]] for t in s] for s in tokens
|
||||
]
|
||||
|
||||
# select the top-ranked sample in each group
|
||||
@ -757,18 +656,9 @@ class DecodingTask:
|
||||
texts: List[str] = [tokenizer.decode(t).strip() for t in tokens]
|
||||
|
||||
sum_logprobs: List[float] = [lp[i] for i, lp in zip(selected, sum_logprobs)]
|
||||
avg_logprobs: List[float] = [
|
||||
lp / (len(t) + 1) for t, lp in zip(tokens, sum_logprobs)
|
||||
]
|
||||
avg_logprobs: List[float] = [lp / (len(t) + 1) for t, lp in zip(tokens, sum_logprobs)]
|
||||
|
||||
fields = (
|
||||
texts,
|
||||
languages,
|
||||
tokens,
|
||||
audio_features,
|
||||
avg_logprobs,
|
||||
no_speech_probs,
|
||||
)
|
||||
fields = (texts, languages, tokens, audio_features, avg_logprobs, no_speech_probs)
|
||||
if len(set(map(len, fields))) != 1:
|
||||
raise RuntimeError(f"inconsistent result lengths: {list(map(len, fields))}")
|
||||
|
||||
@ -783,19 +673,12 @@ class DecodingTask:
|
||||
temperature=self.options.temperature,
|
||||
compression_ratio=compression_ratio(text),
|
||||
)
|
||||
for text, language, tokens, features, avg_logprob, no_speech_prob in zip(
|
||||
*fields
|
||||
)
|
||||
for text, language, tokens, features, avg_logprob, no_speech_prob in zip(*fields)
|
||||
]
|
||||
|
||||
|
||||
@torch.no_grad()
|
||||
def decode(
|
||||
model: "Whisper",
|
||||
mel: Tensor,
|
||||
options: DecodingOptions = DecodingOptions(),
|
||||
**kwargs,
|
||||
) -> Union[DecodingResult, List[DecodingResult]]:
|
||||
def decode(model: "Whisper", mel: Tensor, options: DecodingOptions = DecodingOptions()) -> Union[DecodingResult, List[DecodingResult]]:
|
||||
"""
|
||||
Performs decoding of 30-second audio segment(s), provided as Mel spectrogram(s).
|
||||
|
||||
@ -815,12 +698,13 @@ def decode(
|
||||
result: Union[DecodingResult, List[DecodingResult]]
|
||||
The result(s) of decoding contained in `DecodingResult` dataclass instance(s)
|
||||
"""
|
||||
if single := mel.ndim == 2:
|
||||
single = mel.ndim == 2
|
||||
if single:
|
||||
mel = mel.unsqueeze(0)
|
||||
|
||||
if kwargs:
|
||||
options = replace(options, **kwargs)
|
||||
|
||||
result = DecodingTask(model, options).run(mel)
|
||||
|
||||
if single:
|
||||
result = result[0]
|
||||
|
||||
return result[0] if single else result
|
||||
return result
|
||||
|
||||
133
whisper/model.py
133
whisper/model.py
@ -1,25 +1,15 @@
|
||||
import base64
|
||||
import gzip
|
||||
from contextlib import contextmanager
|
||||
from dataclasses import dataclass
|
||||
from typing import Dict, Iterable, Optional, Tuple
|
||||
from typing import Dict
|
||||
from typing import Iterable, Optional
|
||||
|
||||
import numpy as np
|
||||
import torch
|
||||
import torch.nn.functional as F
|
||||
from torch import Tensor, nn
|
||||
from torch import Tensor
|
||||
from torch import nn
|
||||
|
||||
from .decoding import decode as decode_function
|
||||
from .decoding import detect_language as detect_language_function
|
||||
from .transcribe import transcribe as transcribe_function
|
||||
|
||||
try:
|
||||
from torch.nn.functional import scaled_dot_product_attention
|
||||
|
||||
SDPA_AVAILABLE = True
|
||||
except (ImportError, RuntimeError, OSError):
|
||||
scaled_dot_product_attention = None
|
||||
SDPA_AVAILABLE = False
|
||||
from .decoding import detect_language as detect_language_function, decode as decode_function
|
||||
|
||||
|
||||
@dataclass
|
||||
@ -44,16 +34,12 @@ class LayerNorm(nn.LayerNorm):
|
||||
class Linear(nn.Linear):
|
||||
def forward(self, x: Tensor) -> Tensor:
|
||||
return F.linear(
|
||||
x,
|
||||
self.weight.to(x.dtype),
|
||||
None if self.bias is None else self.bias.to(x.dtype),
|
||||
x, self.weight.to(x.dtype), None if self.bias is None else self.bias.to(x.dtype)
|
||||
)
|
||||
|
||||
|
||||
class Conv1d(nn.Conv1d):
|
||||
def _conv_forward(
|
||||
self, x: Tensor, weight: Tensor, bias: Optional[Tensor]
|
||||
) -> Tensor:
|
||||
def _conv_forward(self, x: Tensor, weight: Tensor, bias: Optional[Tensor]) -> Tensor:
|
||||
return super()._conv_forward(
|
||||
x, weight.to(x.dtype), None if bias is None else bias.to(x.dtype)
|
||||
)
|
||||
@ -68,19 +54,7 @@ def sinusoids(length, channels, max_timescale=10000):
|
||||
return torch.cat([torch.sin(scaled_time), torch.cos(scaled_time)], dim=1)
|
||||
|
||||
|
||||
@contextmanager
|
||||
def disable_sdpa():
|
||||
prev_state = MultiHeadAttention.use_sdpa
|
||||
try:
|
||||
MultiHeadAttention.use_sdpa = False
|
||||
yield
|
||||
finally:
|
||||
MultiHeadAttention.use_sdpa = prev_state
|
||||
|
||||
|
||||
class MultiHeadAttention(nn.Module):
|
||||
use_sdpa = True
|
||||
|
||||
def __init__(self, n_state: int, n_head: int):
|
||||
super().__init__()
|
||||
self.n_head = n_head
|
||||
@ -111,32 +85,20 @@ class MultiHeadAttention(nn.Module):
|
||||
wv, qk = self.qkv_attention(q, k, v, mask)
|
||||
return self.out(wv), qk
|
||||
|
||||
def qkv_attention(
|
||||
self, q: Tensor, k: Tensor, v: Tensor, mask: Optional[Tensor] = None
|
||||
) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
|
||||
def qkv_attention(self, q: Tensor, k: Tensor, v: Tensor, mask: Optional[Tensor] = None):
|
||||
n_batch, n_ctx, n_state = q.shape
|
||||
scale = (n_state // self.n_head) ** -0.25
|
||||
q = q.view(*q.shape[:2], self.n_head, -1).permute(0, 2, 1, 3)
|
||||
k = k.view(*k.shape[:2], self.n_head, -1).permute(0, 2, 1, 3)
|
||||
q = q.view(*q.shape[:2], self.n_head, -1).permute(0, 2, 1, 3) * scale
|
||||
k = k.view(*k.shape[:2], self.n_head, -1).permute(0, 2, 3, 1) * scale
|
||||
v = v.view(*v.shape[:2], self.n_head, -1).permute(0, 2, 1, 3)
|
||||
|
||||
if SDPA_AVAILABLE and MultiHeadAttention.use_sdpa:
|
||||
a = scaled_dot_product_attention(
|
||||
q, k, v, is_causal=mask is not None and n_ctx > 1
|
||||
)
|
||||
out = a.permute(0, 2, 1, 3).flatten(start_dim=2)
|
||||
qk = None
|
||||
else:
|
||||
qk = (q * scale) @ (k * scale).transpose(-1, -2)
|
||||
if mask is not None:
|
||||
qk = qk + mask[:n_ctx, :n_ctx]
|
||||
qk = qk.float()
|
||||
qk = q @ k
|
||||
if mask is not None:
|
||||
qk = qk + mask[:n_ctx, :n_ctx]
|
||||
qk = qk.float()
|
||||
|
||||
w = F.softmax(qk, dim=-1).to(q.dtype)
|
||||
out = (w @ v).permute(0, 2, 1, 3).flatten(start_dim=2)
|
||||
qk = qk.detach()
|
||||
|
||||
return out, qk
|
||||
w = F.softmax(qk, dim=-1).to(q.dtype)
|
||||
return (w @ v).permute(0, 2, 1, 3).flatten(start_dim=2), qk.detach()
|
||||
|
||||
|
||||
class ResidualAttentionBlock(nn.Module):
|
||||
@ -146,15 +108,11 @@ class ResidualAttentionBlock(nn.Module):
|
||||
self.attn = MultiHeadAttention(n_state, n_head)
|
||||
self.attn_ln = LayerNorm(n_state)
|
||||
|
||||
self.cross_attn = (
|
||||
MultiHeadAttention(n_state, n_head) if cross_attention else None
|
||||
)
|
||||
self.cross_attn = MultiHeadAttention(n_state, n_head) if cross_attention else None
|
||||
self.cross_attn_ln = LayerNorm(n_state) if cross_attention else None
|
||||
|
||||
n_mlp = n_state * 4
|
||||
self.mlp = nn.Sequential(
|
||||
Linear(n_state, n_mlp), nn.GELU(), Linear(n_mlp, n_state)
|
||||
)
|
||||
self.mlp = nn.Sequential(Linear(n_state, n_mlp), nn.GELU(), Linear(n_mlp, n_state))
|
||||
self.mlp_ln = LayerNorm(n_state)
|
||||
|
||||
def forward(
|
||||
@ -172,9 +130,7 @@ class ResidualAttentionBlock(nn.Module):
|
||||
|
||||
|
||||
class AudioEncoder(nn.Module):
|
||||
def __init__(
|
||||
self, n_mels: int, n_ctx: int, n_state: int, n_head: int, n_layer: int
|
||||
):
|
||||
def __init__(self, n_mels: int, n_ctx: int, n_state: int, n_head: int, n_layer: int):
|
||||
super().__init__()
|
||||
self.conv1 = Conv1d(n_mels, n_state, kernel_size=3, padding=1)
|
||||
self.conv2 = Conv1d(n_state, n_state, kernel_size=3, stride=2, padding=1)
|
||||
@ -205,19 +161,14 @@ class AudioEncoder(nn.Module):
|
||||
|
||||
|
||||
class TextDecoder(nn.Module):
|
||||
def __init__(
|
||||
self, n_vocab: int, n_ctx: int, n_state: int, n_head: int, n_layer: int
|
||||
):
|
||||
def __init__(self, n_vocab: int, n_ctx: int, n_state: int, n_head: int, n_layer: int):
|
||||
super().__init__()
|
||||
|
||||
self.token_embedding = nn.Embedding(n_vocab, n_state)
|
||||
self.positional_embedding = nn.Parameter(torch.empty(n_ctx, n_state))
|
||||
|
||||
self.blocks: Iterable[ResidualAttentionBlock] = nn.ModuleList(
|
||||
[
|
||||
ResidualAttentionBlock(n_state, n_head, cross_attention=True)
|
||||
for _ in range(n_layer)
|
||||
]
|
||||
[ResidualAttentionBlock(n_state, n_head, cross_attention=True) for _ in range(n_layer)]
|
||||
)
|
||||
self.ln = LayerNorm(n_state)
|
||||
|
||||
@ -228,23 +179,18 @@ class TextDecoder(nn.Module):
|
||||
"""
|
||||
x : torch.LongTensor, shape = (batch_size, <= n_ctx)
|
||||
the text tokens
|
||||
xa : torch.Tensor, shape = (batch_size, n_audio_ctx, n_audio_state)
|
||||
xa : torch.Tensor, shape = (batch_size, n_mels, n_audio_ctx)
|
||||
the encoded audio features to be attended on
|
||||
"""
|
||||
offset = next(iter(kv_cache.values())).shape[1] if kv_cache else 0
|
||||
x = (
|
||||
self.token_embedding(x)
|
||||
+ self.positional_embedding[offset : offset + x.shape[-1]]
|
||||
)
|
||||
x = self.token_embedding(x) + self.positional_embedding[offset : offset + x.shape[-1]]
|
||||
x = x.to(xa.dtype)
|
||||
|
||||
for block in self.blocks:
|
||||
x = block(x, xa, mask=self.mask, kv_cache=kv_cache)
|
||||
|
||||
x = self.ln(x)
|
||||
logits = (
|
||||
x @ torch.transpose(self.token_embedding.weight.to(x.dtype), 0, 1)
|
||||
).float()
|
||||
logits = (x @ torch.transpose(self.token_embedding.weight.to(x.dtype), 0, 1)).float()
|
||||
|
||||
return logits
|
||||
|
||||
@ -267,22 +213,6 @@ class Whisper(nn.Module):
|
||||
self.dims.n_text_head,
|
||||
self.dims.n_text_layer,
|
||||
)
|
||||
# use the last half among the decoder layers for time alignment by default;
|
||||
# to use a specific set of heads, see `set_alignment_heads()` below.
|
||||
all_heads = torch.zeros(
|
||||
self.dims.n_text_layer, self.dims.n_text_head, dtype=torch.bool
|
||||
)
|
||||
all_heads[self.dims.n_text_layer // 2 :] = True
|
||||
self.register_buffer("alignment_heads", all_heads.to_sparse(), persistent=False)
|
||||
|
||||
def set_alignment_heads(self, dump: bytes):
|
||||
array = np.frombuffer(
|
||||
gzip.decompress(base64.b85decode(dump)), dtype=bool
|
||||
).copy()
|
||||
mask = torch.from_numpy(array).reshape(
|
||||
self.dims.n_text_layer, self.dims.n_text_head
|
||||
)
|
||||
self.register_buffer("alignment_heads", mask.to_sparse(), persistent=False)
|
||||
|
||||
def embed_audio(self, mel: torch.Tensor):
|
||||
return self.encoder(mel)
|
||||
@ -290,9 +220,7 @@ class Whisper(nn.Module):
|
||||
def logits(self, tokens: torch.Tensor, audio_features: torch.Tensor):
|
||||
return self.decoder(tokens, audio_features)
|
||||
|
||||
def forward(
|
||||
self, mel: torch.Tensor, tokens: torch.Tensor
|
||||
) -> Dict[str, torch.Tensor]:
|
||||
def forward(self, mel: torch.Tensor, tokens: torch.Tensor) -> Dict[str, torch.Tensor]:
|
||||
return self.decoder(tokens, self.encoder(mel))
|
||||
|
||||
@property
|
||||
@ -301,11 +229,7 @@ class Whisper(nn.Module):
|
||||
|
||||
@property
|
||||
def is_multilingual(self):
|
||||
return self.dims.n_vocab >= 51865
|
||||
|
||||
@property
|
||||
def num_languages(self):
|
||||
return self.dims.n_vocab - 51765 - int(self.is_multilingual)
|
||||
return self.dims.n_vocab == 51865
|
||||
|
||||
def install_kv_cache_hooks(self, cache: Optional[dict] = None):
|
||||
"""
|
||||
@ -325,9 +249,8 @@ class Whisper(nn.Module):
|
||||
hooks = []
|
||||
|
||||
def save_to_cache(module, _, output):
|
||||
if module not in cache or output.shape[1] > self.dims.n_text_ctx:
|
||||
# save as-is, for the first token or cross attention
|
||||
cache[module] = output
|
||||
if module not in cache or output.shape[1] > self.decoder.positional_embedding.shape[0]:
|
||||
cache[module] = output # save as-is, for the first token or cross attention
|
||||
else:
|
||||
cache[module] = torch.cat([cache[module], output], dim=1).detach()
|
||||
return cache[module]
|
||||
|
||||
@ -1,2 +1,2 @@
|
||||
from .basic import BasicTextNormalizer as BasicTextNormalizer
|
||||
from .english import EnglishTextNormalizer as EnglishTextNormalizer
|
||||
from .basic import BasicTextNormalizer
|
||||
from .english import EnglishTextNormalizer
|
||||
|
||||
@ -30,19 +30,15 @@ def remove_symbols_and_diacritics(s: str, keep=""):
|
||||
and drop any diacritics (category 'Mn' and some manual mappings)
|
||||
"""
|
||||
return "".join(
|
||||
(
|
||||
c
|
||||
if c in keep
|
||||
else (
|
||||
ADDITIONAL_DIACRITICS[c]
|
||||
if c in ADDITIONAL_DIACRITICS
|
||||
else (
|
||||
""
|
||||
if unicodedata.category(c) == "Mn"
|
||||
else " " if unicodedata.category(c)[0] in "MSP" else c
|
||||
)
|
||||
)
|
||||
)
|
||||
c
|
||||
if c in keep
|
||||
else ADDITIONAL_DIACRITICS[c]
|
||||
if c in ADDITIONAL_DIACRITICS
|
||||
else ""
|
||||
if unicodedata.category(c) == "Mn"
|
||||
else " "
|
||||
if unicodedata.category(c)[0] in "MSP"
|
||||
else c
|
||||
for c in unicodedata.normalize("NFKD", s)
|
||||
)
|
||||
|
||||
@ -52,16 +48,13 @@ def remove_symbols(s: str):
|
||||
Replace any other markers, symbols, punctuations with a space, keeping diacritics
|
||||
"""
|
||||
return "".join(
|
||||
" " if unicodedata.category(c)[0] in "MSP" else c
|
||||
for c in unicodedata.normalize("NFKC", s)
|
||||
" " if unicodedata.category(c)[0] in "MSP" else c for c in unicodedata.normalize("NFKC", s)
|
||||
)
|
||||
|
||||
|
||||
class BasicTextNormalizer:
|
||||
def __init__(self, remove_diacritics: bool = False, split_letters: bool = False):
|
||||
self.clean = (
|
||||
remove_symbols_and_diacritics if remove_diacritics else remove_symbols
|
||||
)
|
||||
self.clean = remove_symbols_and_diacritics if remove_diacritics else remove_symbols
|
||||
self.split_letters = split_letters
|
||||
|
||||
def __call__(self, s: str):
|
||||
@ -73,8 +66,6 @@ class BasicTextNormalizer:
|
||||
if self.split_letters:
|
||||
s = " ".join(regex.findall(r"\X", s, regex.U))
|
||||
|
||||
s = re.sub(
|
||||
r"\s+", " ", s
|
||||
) # replace any successive whitespace characters with a space
|
||||
s = re.sub(r"\s+", " ", s) # replace any successive whitespace characters with a space
|
||||
|
||||
return s
|
||||
|
||||
@ -1737,5 +1737,6 @@
|
||||
"yoghurt": "yogurt",
|
||||
"yoghurts": "yogurts",
|
||||
"mhm": "hmm",
|
||||
"mm": "hmm",
|
||||
"mmm": "hmm"
|
||||
}
|
||||
@ -84,8 +84,7 @@ class EnglishNumberNormalizer:
|
||||
name.replace("y", "ies"): (value, "s") for name, value in self.tens.items()
|
||||
}
|
||||
self.tens_ordinal = {
|
||||
name.replace("y", "ieth"): (value, "th")
|
||||
for name, value in self.tens.items()
|
||||
name.replace("y", "ieth"): (value, "th") for name, value in self.tens.items()
|
||||
}
|
||||
self.tens_suffixed = {**self.tens_plural, **self.tens_ordinal}
|
||||
|
||||
@ -109,10 +108,7 @@ class EnglishNumberNormalizer:
|
||||
self.multipliers_ordinal = {
|
||||
name + "th": (value, "th") for name, value in self.multipliers.items()
|
||||
}
|
||||
self.multipliers_suffixed = {
|
||||
**self.multipliers_plural,
|
||||
**self.multipliers_ordinal,
|
||||
}
|
||||
self.multipliers_suffixed = {**self.multipliers_plural, **self.multipliers_ordinal}
|
||||
self.decimals = {*self.ones, *self.tens, *self.zeros}
|
||||
|
||||
self.preceding_prefixers = {
|
||||
@ -132,8 +128,7 @@ class EnglishNumberNormalizer:
|
||||
"cents": "¢",
|
||||
}
|
||||
self.prefixes = set(
|
||||
list(self.preceding_prefixers.values())
|
||||
+ list(self.following_prefixers.values())
|
||||
list(self.preceding_prefixers.values()) + list(self.following_prefixers.values())
|
||||
)
|
||||
self.suffixers = {
|
||||
"per": {"cent": "%"},
|
||||
@ -223,9 +218,7 @@ class EnglishNumberNormalizer:
|
||||
if value is None:
|
||||
value = ones
|
||||
elif isinstance(value, str) or prev in self.ones:
|
||||
if (
|
||||
prev in self.tens and ones < 10
|
||||
): # replace the last zero with the digit
|
||||
if prev in self.tens and ones < 10: # replace the last zero with the digit
|
||||
assert value[-1] == "0"
|
||||
value = value[:-1] + str(ones)
|
||||
else:
|
||||
@ -529,14 +522,14 @@ class EnglishTextNormalizer:
|
||||
s = re.sub(r"[<\[][^>\]]*[>\]]", "", s) # remove words between brackets
|
||||
s = re.sub(r"\(([^)]+?)\)", "", s) # remove words between parenthesis
|
||||
s = re.sub(self.ignore_patterns, "", s)
|
||||
s = re.sub(r"\s+'", "'", s) # when there's a space before an apostrophe
|
||||
s = re.sub(r"\s+'", "'", s) # standardize when there's a space before an apostrophe
|
||||
|
||||
for pattern, replacement in self.replacers.items():
|
||||
s = re.sub(pattern, replacement, s)
|
||||
|
||||
s = re.sub(r"(\d),(\d)", r"\1\2", s) # remove commas between digits
|
||||
s = re.sub(r"\.([^0-9]|$)", r" \1", s) # remove periods not followed by numbers
|
||||
s = remove_symbols_and_diacritics(s, keep=".%$¢€£") # keep numeric symbols
|
||||
s = remove_symbols_and_diacritics(s, keep=".%$¢€£") # keep some symbols for numerics
|
||||
|
||||
s = self.standardize_numbers(s)
|
||||
s = self.standardize_spellings(s)
|
||||
@ -545,6 +538,6 @@ class EnglishTextNormalizer:
|
||||
s = re.sub(r"[.$¢€£]([^0-9])", r" \1", s)
|
||||
s = re.sub(r"([^0-9])%", r"\1 ", s)
|
||||
|
||||
s = re.sub(r"\s+", " ", s) # replace any successive whitespaces with a space
|
||||
s = re.sub(r"\s+", " ", s) # replace any successive whitespace characters with a space
|
||||
|
||||
return s
|
||||
|
||||
@ -1,388 +0,0 @@
|
||||
import itertools
|
||||
import subprocess
|
||||
import warnings
|
||||
from dataclasses import dataclass
|
||||
from typing import TYPE_CHECKING, List
|
||||
|
||||
import numba
|
||||
import numpy as np
|
||||
import torch
|
||||
import torch.nn.functional as F
|
||||
|
||||
from .audio import HOP_LENGTH, SAMPLE_RATE, TOKENS_PER_SECOND
|
||||
from .tokenizer import Tokenizer
|
||||
|
||||
if TYPE_CHECKING:
|
||||
from .model import Whisper
|
||||
|
||||
|
||||
def median_filter(x: torch.Tensor, filter_width: int):
|
||||
"""Apply a median filter of width `filter_width` along the last dimension of `x`"""
|
||||
pad_width = filter_width // 2
|
||||
if x.shape[-1] <= pad_width:
|
||||
# F.pad requires the padding width to be smaller than the input dimension
|
||||
return x
|
||||
|
||||
if (ndim := x.ndim) <= 2:
|
||||
# `F.pad` does not support 1D or 2D inputs for reflect padding but supports 3D and 4D
|
||||
x = x[None, None, :]
|
||||
|
||||
assert (
|
||||
filter_width > 0 and filter_width % 2 == 1
|
||||
), "`filter_width` should be an odd number"
|
||||
|
||||
result = None
|
||||
x = F.pad(x, (filter_width // 2, filter_width // 2, 0, 0), mode="reflect")
|
||||
if x.is_cuda:
|
||||
try:
|
||||
from .triton_ops import median_filter_cuda
|
||||
|
||||
result = median_filter_cuda(x, filter_width)
|
||||
except (RuntimeError, subprocess.CalledProcessError):
|
||||
warnings.warn(
|
||||
"Failed to launch Triton kernels, likely due to missing CUDA toolkit; "
|
||||
"falling back to a slower median kernel implementation..."
|
||||
)
|
||||
|
||||
if result is None:
|
||||
# sort() is faster than torch.median (https://github.com/pytorch/pytorch/issues/51450)
|
||||
result = x.unfold(-1, filter_width, 1).sort()[0][..., filter_width // 2]
|
||||
|
||||
if ndim <= 2:
|
||||
result = result[0, 0]
|
||||
|
||||
return result
|
||||
|
||||
|
||||
@numba.jit(nopython=True)
|
||||
def backtrace(trace: np.ndarray):
|
||||
i = trace.shape[0] - 1
|
||||
j = trace.shape[1] - 1
|
||||
trace[0, :] = 2
|
||||
trace[:, 0] = 1
|
||||
|
||||
result = []
|
||||
while i > 0 or j > 0:
|
||||
result.append((i - 1, j - 1))
|
||||
|
||||
if trace[i, j] == 0:
|
||||
i -= 1
|
||||
j -= 1
|
||||
elif trace[i, j] == 1:
|
||||
i -= 1
|
||||
elif trace[i, j] == 2:
|
||||
j -= 1
|
||||
else:
|
||||
raise ValueError("Unexpected trace[i, j]")
|
||||
|
||||
result = np.array(result)
|
||||
return result[::-1, :].T
|
||||
|
||||
|
||||
@numba.jit(nopython=True, parallel=True)
|
||||
def dtw_cpu(x: np.ndarray):
|
||||
N, M = x.shape
|
||||
cost = np.ones((N + 1, M + 1), dtype=np.float32) * np.inf
|
||||
trace = -np.ones((N + 1, M + 1), dtype=np.float32)
|
||||
|
||||
cost[0, 0] = 0
|
||||
for j in range(1, M + 1):
|
||||
for i in range(1, N + 1):
|
||||
c0 = cost[i - 1, j - 1]
|
||||
c1 = cost[i - 1, j]
|
||||
c2 = cost[i, j - 1]
|
||||
|
||||
if c0 < c1 and c0 < c2:
|
||||
c, t = c0, 0
|
||||
elif c1 < c0 and c1 < c2:
|
||||
c, t = c1, 1
|
||||
else:
|
||||
c, t = c2, 2
|
||||
|
||||
cost[i, j] = x[i - 1, j - 1] + c
|
||||
trace[i, j] = t
|
||||
|
||||
return backtrace(trace)
|
||||
|
||||
|
||||
def dtw_cuda(x, BLOCK_SIZE=1024):
|
||||
from .triton_ops import dtw_kernel
|
||||
|
||||
M, N = x.shape
|
||||
assert M < BLOCK_SIZE, f"M should be smaller than {BLOCK_SIZE=}"
|
||||
|
||||
x_skew = (
|
||||
F.pad(x, (0, M + 1), value=np.inf).flatten()[: M * (N + M)].reshape(M, N + M)
|
||||
)
|
||||
x_skew = x_skew.T.contiguous()
|
||||
cost = torch.ones(N + M + 2, M + 2) * np.inf
|
||||
cost[0, 0] = 0
|
||||
cost = cost.to(x.device)
|
||||
trace = torch.zeros_like(cost, dtype=torch.int32)
|
||||
|
||||
dtw_kernel[(1,)](
|
||||
cost,
|
||||
trace,
|
||||
x_skew,
|
||||
x_skew.stride(0),
|
||||
cost.stride(0),
|
||||
trace.stride(0),
|
||||
N,
|
||||
M,
|
||||
BLOCK_SIZE=BLOCK_SIZE,
|
||||
)
|
||||
|
||||
trace = trace.T.flatten()[: (M + 1) * (M + N + 3)].reshape(M + 1, M + N + 3)[
|
||||
:, : N + 1
|
||||
]
|
||||
return backtrace(trace.cpu().numpy())
|
||||
|
||||
|
||||
def dtw(x: torch.Tensor) -> np.ndarray:
|
||||
if x.is_cuda:
|
||||
try:
|
||||
return dtw_cuda(x)
|
||||
except (RuntimeError, subprocess.CalledProcessError):
|
||||
warnings.warn(
|
||||
"Failed to launch Triton kernels, likely due to missing CUDA toolkit; "
|
||||
"falling back to a slower DTW implementation..."
|
||||
)
|
||||
|
||||
return dtw_cpu(x.double().cpu().numpy())
|
||||
|
||||
|
||||
@dataclass
|
||||
class WordTiming:
|
||||
word: str
|
||||
tokens: List[int]
|
||||
start: float
|
||||
end: float
|
||||
probability: float
|
||||
|
||||
|
||||
def find_alignment(
|
||||
model: "Whisper",
|
||||
tokenizer: Tokenizer,
|
||||
text_tokens: List[int],
|
||||
mel: torch.Tensor,
|
||||
num_frames: int,
|
||||
*,
|
||||
medfilt_width: int = 7,
|
||||
qk_scale: float = 1.0,
|
||||
) -> List[WordTiming]:
|
||||
if len(text_tokens) == 0:
|
||||
return []
|
||||
|
||||
tokens = torch.tensor(
|
||||
[
|
||||
*tokenizer.sot_sequence,
|
||||
tokenizer.no_timestamps,
|
||||
*text_tokens,
|
||||
tokenizer.eot,
|
||||
]
|
||||
).to(model.device)
|
||||
|
||||
# install hooks on the cross attention layers to retrieve the attention weights
|
||||
QKs = [None] * model.dims.n_text_layer
|
||||
hooks = [
|
||||
block.cross_attn.register_forward_hook(
|
||||
lambda _, ins, outs, index=i: QKs.__setitem__(index, outs[-1][0])
|
||||
)
|
||||
for i, block in enumerate(model.decoder.blocks)
|
||||
]
|
||||
|
||||
from .model import disable_sdpa
|
||||
|
||||
with torch.no_grad(), disable_sdpa():
|
||||
logits = model(mel.unsqueeze(0), tokens.unsqueeze(0))[0]
|
||||
sampled_logits = logits[len(tokenizer.sot_sequence) :, : tokenizer.eot]
|
||||
token_probs = sampled_logits.softmax(dim=-1)
|
||||
text_token_probs = token_probs[np.arange(len(text_tokens)), text_tokens]
|
||||
text_token_probs = text_token_probs.tolist()
|
||||
|
||||
for hook in hooks:
|
||||
hook.remove()
|
||||
|
||||
# heads * tokens * frames
|
||||
weights = torch.stack([QKs[_l][_h] for _l, _h in model.alignment_heads.indices().T])
|
||||
weights = weights[:, :, : num_frames // 2]
|
||||
weights = (weights * qk_scale).softmax(dim=-1)
|
||||
std, mean = torch.std_mean(weights, dim=-2, keepdim=True, unbiased=False)
|
||||
weights = (weights - mean) / std
|
||||
weights = median_filter(weights, medfilt_width)
|
||||
|
||||
matrix = weights.mean(axis=0)
|
||||
matrix = matrix[len(tokenizer.sot_sequence) : -1]
|
||||
text_indices, time_indices = dtw(-matrix)
|
||||
|
||||
words, word_tokens = tokenizer.split_to_word_tokens(text_tokens + [tokenizer.eot])
|
||||
if len(word_tokens) <= 1:
|
||||
# return on eot only
|
||||
# >>> np.pad([], (1, 0))
|
||||
# array([0.])
|
||||
# This results in crashes when we lookup jump_times with float, like
|
||||
# IndexError: arrays used as indices must be of integer (or boolean) type
|
||||
return []
|
||||
word_boundaries = np.pad(np.cumsum([len(t) for t in word_tokens[:-1]]), (1, 0))
|
||||
|
||||
jumps = np.pad(np.diff(text_indices), (1, 0), constant_values=1).astype(bool)
|
||||
jump_times = time_indices[jumps] / TOKENS_PER_SECOND
|
||||
start_times = jump_times[word_boundaries[:-1]]
|
||||
end_times = jump_times[word_boundaries[1:]]
|
||||
word_probabilities = [
|
||||
np.mean(text_token_probs[i:j])
|
||||
for i, j in zip(word_boundaries[:-1], word_boundaries[1:])
|
||||
]
|
||||
|
||||
return [
|
||||
WordTiming(word, tokens, start, end, probability)
|
||||
for word, tokens, start, end, probability in zip(
|
||||
words, word_tokens, start_times, end_times, word_probabilities
|
||||
)
|
||||
]
|
||||
|
||||
|
||||
def merge_punctuations(alignment: List[WordTiming], prepended: str, appended: str):
|
||||
# merge prepended punctuations
|
||||
i = len(alignment) - 2
|
||||
j = len(alignment) - 1
|
||||
while i >= 0:
|
||||
previous = alignment[i]
|
||||
following = alignment[j]
|
||||
if previous.word.startswith(" ") and previous.word.strip() in prepended:
|
||||
# prepend it to the following word
|
||||
following.word = previous.word + following.word
|
||||
following.tokens = previous.tokens + following.tokens
|
||||
previous.word = ""
|
||||
previous.tokens = []
|
||||
else:
|
||||
j = i
|
||||
i -= 1
|
||||
|
||||
# merge appended punctuations
|
||||
i = 0
|
||||
j = 1
|
||||
while j < len(alignment):
|
||||
previous = alignment[i]
|
||||
following = alignment[j]
|
||||
if not previous.word.endswith(" ") and following.word in appended:
|
||||
# append it to the previous word
|
||||
previous.word = previous.word + following.word
|
||||
previous.tokens = previous.tokens + following.tokens
|
||||
following.word = ""
|
||||
following.tokens = []
|
||||
else:
|
||||
i = j
|
||||
j += 1
|
||||
|
||||
|
||||
def add_word_timestamps(
|
||||
*,
|
||||
segments: List[dict],
|
||||
model: "Whisper",
|
||||
tokenizer: Tokenizer,
|
||||
mel: torch.Tensor,
|
||||
num_frames: int,
|
||||
prepend_punctuations: str = "\"'“¿([{-",
|
||||
append_punctuations: str = "\"'.。,，!！?？:：”)]}、",
|
||||
last_speech_timestamp: float,
|
||||
**kwargs,
|
||||
):
|
||||
if len(segments) == 0:
|
||||
return
|
||||
|
||||
text_tokens_per_segment = [
|
||||
[token for token in segment["tokens"] if token < tokenizer.eot]
|
||||
for segment in segments
|
||||
]
|
||||
|
||||
text_tokens = list(itertools.chain.from_iterable(text_tokens_per_segment))
|
||||
alignment = find_alignment(model, tokenizer, text_tokens, mel, num_frames, **kwargs)
|
||||
word_durations = np.array([t.end - t.start for t in alignment])
|
||||
word_durations = word_durations[word_durations.nonzero()]
|
||||
median_duration = np.median(word_durations) if len(word_durations) > 0 else 0.0
|
||||
median_duration = min(0.7, float(median_duration))
|
||||
max_duration = median_duration * 2
|
||||
|
||||
# hack: truncate long words at sentence boundaries.
|
||||
# a better segmentation algorithm based on VAD should be able to replace this.
|
||||
if len(word_durations) > 0:
|
||||
sentence_end_marks = ".。!！?？"
|
||||
# ensure words at sentence boundaries are not longer than twice the median word duration.
|
||||
for i in range(1, len(alignment)):
|
||||
if alignment[i].end - alignment[i].start > max_duration:
|
||||
if alignment[i].word in sentence_end_marks:
|
||||
alignment[i].end = alignment[i].start + max_duration
|
||||
elif alignment[i - 1].word in sentence_end_marks:
|
||||
alignment[i].start = alignment[i].end - max_duration
|
||||
|
||||
merge_punctuations(alignment, prepend_punctuations, append_punctuations)
|
||||
|
||||
time_offset = segments[0]["seek"] * HOP_LENGTH / SAMPLE_RATE
|
||||
word_index = 0
|
||||
|
||||
for segment, text_tokens in zip(segments, text_tokens_per_segment):
|
||||
saved_tokens = 0
|
||||
words = []
|
||||
|
||||
while word_index < len(alignment) and saved_tokens < len(text_tokens):
|
||||
timing = alignment[word_index]
|
||||
|
||||
if timing.word:
|
||||
words.append(
|
||||
dict(
|
||||
word=timing.word,
|
||||
start=round(time_offset + timing.start, 2),
|
||||
end=round(time_offset + timing.end, 2),
|
||||
probability=timing.probability,
|
||||
)
|
||||
)
|
||||
|
||||
saved_tokens += len(timing.tokens)
|
||||
word_index += 1
|
||||
|
||||
# hack: truncate long words at segment boundaries.
|
||||
# a better segmentation algorithm based on VAD should be able to replace this.
|
||||
if len(words) > 0:
|
||||
# ensure the first and second word after a pause is not longer than
|
||||
# twice the median word duration.
|
||||
if words[0]["end"] - last_speech_timestamp > median_duration * 4 and (
|
||||
words[0]["end"] - words[0]["start"] > max_duration
|
||||
or (
|
||||
len(words) > 1
|
||||
and words[1]["end"] - words[0]["start"] > max_duration * 2
|
||||
)
|
||||
):
|
||||
if (
|
||||
len(words) > 1
|
||||
and words[1]["end"] - words[1]["start"] > max_duration
|
||||
):
|
||||
boundary = max(words[1]["end"] / 2, words[1]["end"] - max_duration)
|
||||
words[0]["end"] = words[1]["start"] = boundary
|
||||
words[0]["start"] = max(0, words[0]["end"] - max_duration)
|
||||
|
||||
# prefer the segment-level start timestamp if the first word is too long.
|
||||
if (
|
||||
segment["start"] < words[0]["end"]
|
||||
and segment["start"] - 0.5 > words[0]["start"]
|
||||
):
|
||||
words[0]["start"] = max(
|
||||
0, min(words[0]["end"] - median_duration, segment["start"])
|
||||
)
|
||||
else:
|
||||
segment["start"] = words[0]["start"]
|
||||
|
||||
# prefer the segment-level end timestamp if the last word is too long.
|
||||
if (
|
||||
segment["end"] > words[-1]["start"]
|
||||
and segment["end"] + 0.5 < words[-1]["end"]
|
||||
):
|
||||
words[-1]["end"] = max(
|
||||
words[-1]["start"] + median_duration, segment["end"]
|
||||
)
|
||||
else:
|
||||
segment["end"] = words[-1]["end"]
|
||||
|
||||
last_speech_timestamp = segment["end"]
|
||||
|
||||
segment["words"] = words
|
||||
@ -1,11 +1,11 @@
|
||||
import base64
|
||||
import os
|
||||
import string
|
||||
from dataclasses import dataclass, field
|
||||
from functools import cached_property, lru_cache
|
||||
from typing import Dict, List, Optional, Tuple
|
||||
from dataclasses import dataclass
|
||||
from functools import lru_cache
|
||||
from typing import List, Optional, Tuple, Union
|
||||
|
||||
import tiktoken
|
||||
import numpy as np
|
||||
import torch
|
||||
from transformers import GPT2TokenizerFast
|
||||
|
||||
LANGUAGES = {
|
||||
"en": "english",
|
||||
@ -107,7 +107,6 @@ LANGUAGES = {
|
||||
"ba": "bashkir",
|
||||
"jw": "javanese",
|
||||
"su": "sundanese",
|
||||
"yue": "cantonese",
|
||||
}
|
||||
|
||||
# language code lookup by name, with a few language aliases
|
||||
@ -124,121 +123,117 @@ TO_LANGUAGE_CODE = {
|
||||
"moldovan": "ro",
|
||||
"sinhalese": "si",
|
||||
"castilian": "es",
|
||||
"mandarin": "zh",
|
||||
}
|
||||
|
||||
|
||||
@dataclass
|
||||
@dataclass(frozen=True)
|
||||
class Tokenizer:
|
||||
"""A thin wrapper around `tiktoken` providing quick access to special tokens"""
|
||||
"""A thin wrapper around `GPT2TokenizerFast` providing quick access to special tokens"""
|
||||
|
||||
encoding: tiktoken.Encoding
|
||||
num_languages: int
|
||||
language: Optional[str] = None
|
||||
task: Optional[str] = None
|
||||
sot_sequence: Tuple[int] = ()
|
||||
special_tokens: Dict[str, int] = field(default_factory=dict)
|
||||
|
||||
def __post_init__(self):
|
||||
for special in self.encoding.special_tokens_set:
|
||||
special_token = self.encoding.encode_single_token(special)
|
||||
self.special_tokens[special] = special_token
|
||||
|
||||
sot: int = self.special_tokens["<|startoftranscript|>"]
|
||||
translate: int = self.special_tokens["<|translate|>"]
|
||||
transcribe: int = self.special_tokens["<|transcribe|>"]
|
||||
|
||||
langs = tuple(LANGUAGES.keys())[: self.num_languages]
|
||||
sot_sequence = [sot]
|
||||
if self.language is not None:
|
||||
sot_sequence.append(sot + 1 + langs.index(self.language))
|
||||
if self.task is not None:
|
||||
task_token: int = transcribe if self.task == "transcribe" else translate
|
||||
sot_sequence.append(task_token)
|
||||
|
||||
self.sot_sequence = tuple(sot_sequence)
|
||||
tokenizer: "GPT2TokenizerFast"
|
||||
language: Optional[str]
|
||||
sot_sequence: Tuple[int]
|
||||
|
||||
def encode(self, text, **kwargs):
|
||||
return self.encoding.encode(text, **kwargs)
|
||||
return self.tokenizer.encode(text, **kwargs)
|
||||
|
||||
def decode(self, token_ids: List[int], **kwargs) -> str:
|
||||
token_ids = [t for t in token_ids if t < self.timestamp_begin]
|
||||
return self.encoding.decode(token_ids, **kwargs)
|
||||
def decode(self, token_ids: Union[int, List[int], np.ndarray, torch.Tensor], **kwargs):
|
||||
return self.tokenizer.decode(token_ids, **kwargs)
|
||||
|
||||
def decode_with_timestamps(self, token_ids: List[int], **kwargs) -> str:
|
||||
def decode_with_timestamps(self, tokens) -> str:
|
||||
"""
|
||||
Timestamp tokens are above other special tokens' id range and are ignored by `decode()`.
|
||||
Timestamp tokens are above the special tokens' id range and are ignored by `decode()`.
|
||||
This method decodes given tokens with timestamps tokens annotated, e.g. "<|1.08|>".
|
||||
"""
|
||||
return self.encoding.decode(token_ids, **kwargs)
|
||||
outputs = [[]]
|
||||
for token in tokens:
|
||||
if token >= self.timestamp_begin:
|
||||
timestamp = f"<|{(token - self.timestamp_begin) * 0.02:.2f}|>"
|
||||
outputs.append(timestamp)
|
||||
outputs.append([])
|
||||
else:
|
||||
outputs[-1].append(token)
|
||||
outputs = [s if isinstance(s, str) else self.tokenizer.decode(s) for s in outputs]
|
||||
return "".join(outputs)
|
||||
|
||||
@cached_property
|
||||
@property
|
||||
@lru_cache()
|
||||
def eot(self) -> int:
|
||||
return self.encoding.eot_token
|
||||
return self.tokenizer.eos_token_id
|
||||
|
||||
@cached_property
|
||||
def transcribe(self) -> int:
|
||||
return self.special_tokens["<|transcribe|>"]
|
||||
|
||||
@cached_property
|
||||
def translate(self) -> int:
|
||||
return self.special_tokens["<|translate|>"]
|
||||
|
||||
@cached_property
|
||||
@property
|
||||
@lru_cache()
|
||||
def sot(self) -> int:
|
||||
return self.special_tokens["<|startoftranscript|>"]
|
||||
return self._get_single_token_id("<|startoftranscript|>")
|
||||
|
||||
@cached_property
|
||||
@property
|
||||
@lru_cache()
|
||||
def sot_lm(self) -> int:
|
||||
return self.special_tokens["<|startoflm|>"]
|
||||
return self._get_single_token_id("<|startoflm|>")
|
||||
|
||||
@cached_property
|
||||
@property
|
||||
@lru_cache()
|
||||
def sot_prev(self) -> int:
|
||||
return self.special_tokens["<|startofprev|>"]
|
||||
return self._get_single_token_id("<|startofprev|>")
|
||||
|
||||
@cached_property
|
||||
@property
|
||||
@lru_cache()
|
||||
def no_speech(self) -> int:
|
||||
return self.special_tokens["<|nospeech|>"]
|
||||
return self._get_single_token_id("<|nospeech|>")
|
||||
|
||||
@cached_property
|
||||
@property
|
||||
@lru_cache()
|
||||
def no_timestamps(self) -> int:
|
||||
return self.special_tokens["<|notimestamps|>"]
|
||||
return self._get_single_token_id("<|notimestamps|>")
|
||||
|
||||
@cached_property
|
||||
@property
|
||||
@lru_cache()
|
||||
def timestamp_begin(self) -> int:
|
||||
return self.special_tokens["<|0.00|>"]
|
||||
return self.tokenizer.all_special_ids[-1] + 1
|
||||
|
||||
@cached_property
|
||||
@property
|
||||
@lru_cache()
|
||||
def language_token(self) -> int:
|
||||
"""Returns the token id corresponding to the value of the `language` field"""
|
||||
if self.language is None:
|
||||
raise ValueError("This tokenizer does not have language token configured")
|
||||
raise ValueError(f"This tokenizer does not have language token configured")
|
||||
|
||||
return self.to_language_token(self.language)
|
||||
additional_tokens = dict(
|
||||
zip(
|
||||
self.tokenizer.additional_special_tokens,
|
||||
self.tokenizer.additional_special_tokens_ids,
|
||||
)
|
||||
)
|
||||
candidate = f"<|{self.language}|>"
|
||||
if candidate in additional_tokens:
|
||||
return additional_tokens[candidate]
|
||||
|
||||
def to_language_token(self, language):
|
||||
if token := self.special_tokens.get(f"<|{language}|>", None):
|
||||
return token
|
||||
raise KeyError(f"Language {self.language} not found in tokenizer.")
|
||||
|
||||
raise KeyError(f"Language {language} not found in tokenizer.")
|
||||
|
||||
@cached_property
|
||||
@property
|
||||
@lru_cache()
|
||||
def all_language_tokens(self) -> Tuple[int]:
|
||||
result = []
|
||||
for token, token_id in self.special_tokens.items():
|
||||
for token, token_id in zip(
|
||||
self.tokenizer.additional_special_tokens,
|
||||
self.tokenizer.additional_special_tokens_ids,
|
||||
):
|
||||
if token.strip("<|>") in LANGUAGES:
|
||||
result.append(token_id)
|
||||
return tuple(result)[: self.num_languages]
|
||||
return tuple(result)
|
||||
|
||||
@cached_property
|
||||
@property
|
||||
@lru_cache()
|
||||
def all_language_codes(self) -> Tuple[str]:
|
||||
return tuple(self.decode([_l]).strip("<|>") for _l in self.all_language_tokens)
|
||||
return tuple(self.decode([l]).strip("<|>") for l in self.all_language_tokens)
|
||||
|
||||
@cached_property
|
||||
@property
|
||||
@lru_cache()
|
||||
def sot_sequence_including_notimestamps(self) -> Tuple[int]:
|
||||
return tuple(list(self.sot_sequence) + [self.no_timestamps])
|
||||
|
||||
@cached_property
|
||||
@property
|
||||
@lru_cache()
|
||||
def non_speech_tokens(self) -> Tuple[int]:
|
||||
"""
|
||||
Returns the list of tokens to suppress in order to avoid any speaker tags or non-speech
|
||||
@ -250,10 +245,8 @@ class Tokenizer:
|
||||
|
||||
keeping basic punctuations like commas, periods, question marks, exclamation points, etc.
|
||||
"""
|
||||
symbols = list('"#()*+/:;<=>@[\\]^_`{|}~「」『』')
|
||||
symbols += (
|
||||
"<< >> <<< >>> -- --- -( -[ (' (\" (( )) ((( ))) [[ ]] {{ }} ♪♪ ♪♪♪".split()
|
||||
)
|
||||
symbols = list("\"#()*+/:;<=>@[\\]^_`{|}~「」『』")
|
||||
symbols += "<< >> <<< >>> -- --- -( -[ (' (\" (( )) ((( ))) [[ ]] {{ }} ♪♪ ♪♪♪".split()
|
||||
|
||||
# symbols that may be a single token or multiple tokens depending on the tokenizer.
|
||||
# In case they're multiple tokens, suppress the first token, which is safe because:
|
||||
@ -263,113 +256,47 @@ class Tokenizer:
|
||||
assert all(0x2640 <= ord(c) <= 0x267F for c in miscellaneous)
|
||||
|
||||
# allow hyphens "-" and single quotes "'" between words, but not at the beginning of a word
|
||||
result = {self.encoding.encode(" -")[0], self.encoding.encode(" '")[0]}
|
||||
result = {self.tokenizer.encode(" -")[0], self.tokenizer.encode(" '")[0]}
|
||||
for symbol in symbols + list(miscellaneous):
|
||||
for tokens in [
|
||||
self.encoding.encode(symbol),
|
||||
self.encoding.encode(" " + symbol),
|
||||
]:
|
||||
for tokens in [self.tokenizer.encode(symbol), self.tokenizer.encode(" " + symbol)]:
|
||||
if len(tokens) == 1 or symbol in miscellaneous:
|
||||
result.add(tokens[0])
|
||||
|
||||
return tuple(sorted(result))
|
||||
|
||||
def split_to_word_tokens(self, tokens: List[int]):
|
||||
if self.language in {"zh", "ja", "th", "lo", "my", "yue"}:
|
||||
# These languages don't typically use spaces, so it is difficult to split words
|
||||
# without morpheme analysis. Here, we instead split words at any
|
||||
# position where the tokens are decoded as valid unicode points
|
||||
return self.split_tokens_on_unicode(tokens)
|
||||
|
||||
return self.split_tokens_on_spaces(tokens)
|
||||
|
||||
def split_tokens_on_unicode(self, tokens: List[int]):
|
||||
decoded_full = self.decode_with_timestamps(tokens)
|
||||
replacement_char = "\ufffd"
|
||||
|
||||
words = []
|
||||
word_tokens = []
|
||||
current_tokens = []
|
||||
unicode_offset = 0
|
||||
|
||||
for token in tokens:
|
||||
current_tokens.append(token)
|
||||
decoded = self.decode_with_timestamps(current_tokens)
|
||||
|
||||
if (
|
||||
replacement_char not in decoded
|
||||
or decoded_full[unicode_offset + decoded.index(replacement_char)]
|
||||
== replacement_char
|
||||
):
|
||||
words.append(decoded)
|
||||
word_tokens.append(current_tokens)
|
||||
current_tokens = []
|
||||
unicode_offset += len(decoded)
|
||||
|
||||
return words, word_tokens
|
||||
|
||||
def split_tokens_on_spaces(self, tokens: List[int]):
|
||||
subwords, subword_tokens_list = self.split_tokens_on_unicode(tokens)
|
||||
words = []
|
||||
word_tokens = []
|
||||
|
||||
for subword, subword_tokens in zip(subwords, subword_tokens_list):
|
||||
special = subword_tokens[0] >= self.eot
|
||||
with_space = subword.startswith(" ")
|
||||
punctuation = subword.strip() in string.punctuation
|
||||
if special or with_space or punctuation or len(words) == 0:
|
||||
words.append(subword)
|
||||
word_tokens.append(subword_tokens)
|
||||
else:
|
||||
words[-1] = words[-1] + subword
|
||||
word_tokens[-1].extend(subword_tokens)
|
||||
|
||||
return words, word_tokens
|
||||
def _get_single_token_id(self, text) -> int:
|
||||
tokens = self.tokenizer.encode(text)
|
||||
assert len(tokens) == 1, f"{text} is not encoded as a single token"
|
||||
return tokens[0]
|
||||
|
||||
|
||||
@lru_cache(maxsize=None)
|
||||
def get_encoding(name: str = "gpt2", num_languages: int = 99):
|
||||
vocab_path = os.path.join(os.path.dirname(__file__), "assets", f"{name}.tiktoken")
|
||||
ranks = {
|
||||
base64.b64decode(token): int(rank)
|
||||
for token, rank in (line.split() for line in open(vocab_path) if line)
|
||||
}
|
||||
n_vocab = len(ranks)
|
||||
special_tokens = {}
|
||||
def build_tokenizer(name: str = "gpt2"):
|
||||
os.environ["TOKENIZERS_PARALLELISM"] = "false"
|
||||
path = os.path.join(os.path.dirname(__file__), "assets", name)
|
||||
tokenizer = GPT2TokenizerFast.from_pretrained(path)
|
||||
|
||||
specials = [
|
||||
"<|endoftext|>",
|
||||
"<|startoftranscript|>",
|
||||
*[f"<|{lang}|>" for lang in list(LANGUAGES.keys())[:num_languages]],
|
||||
*[f"<|{lang}|>" for lang in LANGUAGES.keys()],
|
||||
"<|translate|>",
|
||||
"<|transcribe|>",
|
||||
"<|startoflm|>",
|
||||
"<|startofprev|>",
|
||||
"<|nospeech|>",
|
||||
"<|notimestamps|>",
|
||||
*[f"<|{i * 0.02:.2f}|>" for i in range(1501)],
|
||||
]
|
||||
|
||||
for token in specials:
|
||||
special_tokens[token] = n_vocab
|
||||
n_vocab += 1
|
||||
|
||||
return tiktoken.Encoding(
|
||||
name=os.path.basename(vocab_path),
|
||||
explicit_n_vocab=n_vocab,
|
||||
pat_str=r"""'s|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+""",
|
||||
mergeable_ranks=ranks,
|
||||
special_tokens=special_tokens,
|
||||
)
|
||||
tokenizer.add_special_tokens(dict(additional_special_tokens=specials))
|
||||
return tokenizer
|
||||
|
||||
|
||||
@lru_cache(maxsize=None)
|
||||
def get_tokenizer(
|
||||
multilingual: bool,
|
||||
*,
|
||||
num_languages: int = 99,
|
||||
language: Optional[str] = None,
|
||||
task: Optional[str] = None, # Literal["transcribe", "translate", None]
|
||||
language: Optional[str] = None,
|
||||
) -> Tokenizer:
|
||||
if language is not None:
|
||||
language = language.lower()
|
||||
@ -380,16 +307,25 @@ def get_tokenizer(
|
||||
raise ValueError(f"Unsupported language: {language}")
|
||||
|
||||
if multilingual:
|
||||
encoding_name = "multilingual"
|
||||
language = language or "en"
|
||||
tokenizer_name = "multilingual"
|
||||
task = task or "transcribe"
|
||||
language = language or "en"
|
||||
else:
|
||||
encoding_name = "gpt2"
|
||||
language = None
|
||||
tokenizer_name = "gpt2"
|
||||
task = None
|
||||
language = None
|
||||
|
||||
encoding = get_encoding(name=encoding_name, num_languages=num_languages)
|
||||
tokenizer = build_tokenizer(name=tokenizer_name)
|
||||
all_special_ids: List[int] = tokenizer.all_special_ids
|
||||
sot: int = all_special_ids[1]
|
||||
translate: int = all_special_ids[-6]
|
||||
transcribe: int = all_special_ids[-5]
|
||||
|
||||
return Tokenizer(
|
||||
encoding=encoding, num_languages=num_languages, language=language, task=task
|
||||
)
|
||||
langs = tuple(LANGUAGES.keys())
|
||||
sot_sequence = [sot]
|
||||
if language is not None:
|
||||
sot_sequence.append(sot + 1 + langs.index(language))
|
||||
if task is not None:
|
||||
sot_sequence.append(transcribe if task == "transcribe" else translate)
|
||||
|
||||
return Tokenizer(tokenizer=tokenizer, language=language, sot_sequence=tuple(sot_sequence))
|
||||
|
||||
@ -1,35 +1,16 @@
|
||||
import argparse
|
||||
import os
|
||||
import traceback
|
||||
import warnings
|
||||
from typing import TYPE_CHECKING, List, Optional, Tuple, Union
|
||||
from typing import List, Optional, Tuple, Union, TYPE_CHECKING
|
||||
|
||||
import numpy as np
|
||||
import torch
|
||||
import tqdm
|
||||
|
||||
from .audio import (
|
||||
FRAMES_PER_SECOND,
|
||||
HOP_LENGTH,
|
||||
N_FRAMES,
|
||||
N_SAMPLES,
|
||||
SAMPLE_RATE,
|
||||
log_mel_spectrogram,
|
||||
pad_or_trim,
|
||||
)
|
||||
from .audio import SAMPLE_RATE, N_FRAMES, HOP_LENGTH, pad_or_trim, log_mel_spectrogram
|
||||
from .decoding import DecodingOptions, DecodingResult
|
||||
from .timing import add_word_timestamps
|
||||
from .tokenizer import LANGUAGES, TO_LANGUAGE_CODE, get_tokenizer
|
||||
from .utils import (
|
||||
exact_div,
|
||||
format_timestamp,
|
||||
get_end,
|
||||
get_writer,
|
||||
make_safe,
|
||||
optional_float,
|
||||
optional_int,
|
||||
str2bool,
|
||||
)
|
||||
from .utils import exact_div, format_timestamp, optional_int, optional_float, str2bool, write_txt, write_vtt, write_srt
|
||||
|
||||
if TYPE_CHECKING:
|
||||
from .model import Whisper
|
||||
@ -45,13 +26,6 @@ def transcribe(
|
||||
logprob_threshold: Optional[float] = -1.0,
|
||||
no_speech_threshold: Optional[float] = 0.6,
|
||||
condition_on_previous_text: bool = True,
|
||||
initial_prompt: Optional[str] = None,
|
||||
carry_initial_prompt: bool = False,
|
||||
word_timestamps: bool = False,
|
||||
prepend_punctuations: str = "\"'“¿([{-",
|
||||
append_punctuations: str = "\"'.。,，!！?？:：”)]}、",
|
||||
clip_timestamps: Union[str, List[float]] = "0",
|
||||
hallucination_silence_threshold: Optional[float] = None,
|
||||
**decode_options,
|
||||
):
|
||||
"""
|
||||
@ -88,37 +62,9 @@ def transcribe(
|
||||
disabling may make the text inconsistent across windows, but the model becomes less prone to
|
||||
getting stuck in a failure loop, such as repetition looping or timestamps going out of sync.
|
||||
|
||||
word_timestamps: bool
|
||||
Extract word-level timestamps using the cross-attention pattern and dynamic time warping,
|
||||
and include the timestamps for each word in each segment.
|
||||
|
||||
prepend_punctuations: str
|
||||
If word_timestamps is True, merge these punctuation symbols with the next word
|
||||
|
||||
append_punctuations: str
|
||||
If word_timestamps is True, merge these punctuation symbols with the previous word
|
||||
|
||||
initial_prompt: Optional[str]
|
||||
Optional text to provide as a prompt for the first window. This can be used to provide, or
|
||||
"prompt-engineer" a context for transcription, e.g. custom vocabularies or proper nouns
|
||||
to make it more likely to predict those word correctly.
|
||||
|
||||
carry_initial_prompt: bool
|
||||
If carry_initial_prompt is True, `initial_prompt` is prepended to the prompt of each internal
|
||||
`decode()` call. If there is not enough context space at the start of the prompt, it is
|
||||
left-sliced to make space.
|
||||
|
||||
decode_options: dict
|
||||
Keyword arguments to construct `DecodingOptions` instances
|
||||
|
||||
clip_timestamps: Union[str, List[float]]
|
||||
Comma-separated list start,end,start,end,... timestamps (in seconds) of clips to process.
|
||||
The last end timestamp defaults to the end of the file.
|
||||
|
||||
hallucination_silence_threshold: Optional[float]
|
||||
When word_timestamps is True, skip silent periods longer than this threshold (in seconds)
|
||||
when a possible hallucination is detected
|
||||
|
||||
Returns
|
||||
-------
|
||||
A dictionary containing the resulting text ("text") and segment-level details ("segments"), and
|
||||
@ -135,56 +81,26 @@ def transcribe(
|
||||
if dtype == torch.float32:
|
||||
decode_options["fp16"] = False
|
||||
|
||||
# Pad 30-seconds of silence to the input audio, for slicing
|
||||
mel = log_mel_spectrogram(audio, model.dims.n_mels, padding=N_SAMPLES)
|
||||
content_frames = mel.shape[-1] - N_FRAMES
|
||||
content_duration = float(content_frames * HOP_LENGTH / SAMPLE_RATE)
|
||||
mel = log_mel_spectrogram(audio)
|
||||
|
||||
if decode_options.get("language", None) is None:
|
||||
if not model.is_multilingual:
|
||||
decode_options["language"] = "en"
|
||||
else:
|
||||
if verbose:
|
||||
print(
|
||||
"Detecting language using up to the first 30 seconds. Use `--language` to specify the language"
|
||||
)
|
||||
mel_segment = pad_or_trim(mel, N_FRAMES).to(model.device).to(dtype)
|
||||
_, probs = model.detect_language(mel_segment)
|
||||
print("Detecting language using up to the first 30 seconds. Use `--language` to specify the language")
|
||||
segment = pad_or_trim(mel, N_FRAMES).to(model.device).to(dtype)
|
||||
_, probs = model.detect_language(segment)
|
||||
decode_options["language"] = max(probs, key=probs.get)
|
||||
if verbose is not None:
|
||||
print(
|
||||
f"Detected language: {LANGUAGES[decode_options['language']].title()}"
|
||||
)
|
||||
print(f"Detected language: {LANGUAGES[decode_options['language']].title()}")
|
||||
|
||||
language: str = decode_options["language"]
|
||||
task: str = decode_options.get("task", "transcribe")
|
||||
tokenizer = get_tokenizer(
|
||||
model.is_multilingual,
|
||||
num_languages=model.num_languages,
|
||||
language=language,
|
||||
task=task,
|
||||
)
|
||||
|
||||
if isinstance(clip_timestamps, str):
|
||||
clip_timestamps = [
|
||||
float(ts) for ts in (clip_timestamps.split(",") if clip_timestamps else [])
|
||||
]
|
||||
seek_points: List[int] = [round(ts * FRAMES_PER_SECOND) for ts in clip_timestamps]
|
||||
if len(seek_points) == 0:
|
||||
seek_points.append(0)
|
||||
if len(seek_points) % 2 == 1:
|
||||
seek_points.append(content_frames)
|
||||
seek_clips: List[Tuple[int, int]] = list(zip(seek_points[::2], seek_points[1::2]))
|
||||
|
||||
punctuation = "\"'“¿([{-\"'.。,，!！?？:：”)]}、"
|
||||
|
||||
if word_timestamps and task == "translate":
|
||||
warnings.warn("Word-level timestamps on translations may not be reliable.")
|
||||
language = decode_options["language"]
|
||||
task = decode_options.get("task", "transcribe")
|
||||
tokenizer = get_tokenizer(model.is_multilingual, language=language, task=task)
|
||||
|
||||
def decode_with_fallback(segment: torch.Tensor) -> DecodingResult:
|
||||
temperatures = (
|
||||
[temperature] if isinstance(temperature, (int, float)) else temperature
|
||||
)
|
||||
temperatures = [temperature] if isinstance(temperature, (int, float)) else temperature
|
||||
decode_result = None
|
||||
|
||||
for t in temperatures:
|
||||
@ -201,30 +117,17 @@ def transcribe(
|
||||
decode_result = model.decode(segment, options)
|
||||
|
||||
needs_fallback = False
|
||||
if (
|
||||
compression_ratio_threshold is not None
|
||||
and decode_result.compression_ratio > compression_ratio_threshold
|
||||
):
|
||||
if compression_ratio_threshold is not None and decode_result.compression_ratio > compression_ratio_threshold:
|
||||
needs_fallback = True # too repetitive
|
||||
if (
|
||||
logprob_threshold is not None
|
||||
and decode_result.avg_logprob < logprob_threshold
|
||||
):
|
||||
if logprob_threshold is not None and decode_result.avg_logprob < logprob_threshold:
|
||||
needs_fallback = True # average log probability is too low
|
||||
if (
|
||||
no_speech_threshold is not None
|
||||
and decode_result.no_speech_prob > no_speech_threshold
|
||||
and logprob_threshold is not None
|
||||
and decode_result.avg_logprob < logprob_threshold
|
||||
):
|
||||
needs_fallback = False # silence
|
||||
|
||||
if not needs_fallback:
|
||||
break
|
||||
|
||||
return decode_result
|
||||
|
||||
clip_idx = 0
|
||||
seek = seek_clips[clip_idx][0]
|
||||
seek = 0
|
||||
input_stride = exact_div(
|
||||
N_FRAMES, model.dims.n_audio_ctx
|
||||
) # mel frames per output token: 2
|
||||
@ -235,303 +138,123 @@ def transcribe(
|
||||
all_segments = []
|
||||
prompt_reset_since = 0
|
||||
|
||||
remaining_prompt_length = model.dims.n_text_ctx // 2 - 1
|
||||
if initial_prompt is not None:
|
||||
initial_prompt_tokens = tokenizer.encode(" " + initial_prompt.strip())
|
||||
all_tokens.extend(initial_prompt_tokens)
|
||||
remaining_prompt_length -= len(initial_prompt_tokens)
|
||||
else:
|
||||
initial_prompt_tokens = []
|
||||
initial_prompt = decode_options.pop("initial_prompt", None) or []
|
||||
if initial_prompt:
|
||||
initial_prompt = tokenizer.encode(" " + initial_prompt.strip())
|
||||
all_tokens.extend(initial_prompt)
|
||||
|
||||
def new_segment(
|
||||
*, start: float, end: float, tokens: torch.Tensor, result: DecodingResult
|
||||
def add_segment(
|
||||
*, start: float, end: float, text_tokens: torch.Tensor, result: DecodingResult
|
||||
):
|
||||
tokens = tokens.tolist()
|
||||
text_tokens = [token for token in tokens if token < tokenizer.eot]
|
||||
return {
|
||||
"seek": seek,
|
||||
"start": start,
|
||||
"end": end,
|
||||
"text": tokenizer.decode(text_tokens),
|
||||
"tokens": tokens,
|
||||
"temperature": result.temperature,
|
||||
"avg_logprob": result.avg_logprob,
|
||||
"compression_ratio": result.compression_ratio,
|
||||
"no_speech_prob": result.no_speech_prob,
|
||||
}
|
||||
text = tokenizer.decode([token for token in text_tokens if token < tokenizer.eot])
|
||||
if len(text.strip()) == 0: # skip empty text output
|
||||
return
|
||||
|
||||
# show the progress bar when verbose is False (if True, transcribed text will be printed)
|
||||
with tqdm.tqdm(
|
||||
total=content_frames, unit="frames", disable=verbose is not False
|
||||
) as pbar:
|
||||
last_speech_timestamp = 0.0
|
||||
# NOTE: This loop is obscurely flattened to make the diff readable.
|
||||
# A later commit should turn this into a simpler nested loop.
|
||||
# for seek_clip_start, seek_clip_end in seek_clips:
|
||||
# while seek < seek_clip_end
|
||||
while clip_idx < len(seek_clips):
|
||||
seek_clip_start, seek_clip_end = seek_clips[clip_idx]
|
||||
if seek < seek_clip_start:
|
||||
seek = seek_clip_start
|
||||
if seek >= seek_clip_end:
|
||||
clip_idx += 1
|
||||
if clip_idx < len(seek_clips):
|
||||
seek = seek_clips[clip_idx][0]
|
||||
continue
|
||||
time_offset = float(seek * HOP_LENGTH / SAMPLE_RATE)
|
||||
window_end_time = float((seek + N_FRAMES) * HOP_LENGTH / SAMPLE_RATE)
|
||||
segment_size = min(N_FRAMES, content_frames - seek, seek_clip_end - seek)
|
||||
mel_segment = mel[:, seek : seek + segment_size]
|
||||
segment_duration = segment_size * HOP_LENGTH / SAMPLE_RATE
|
||||
mel_segment = pad_or_trim(mel_segment, N_FRAMES).to(model.device).to(dtype)
|
||||
all_segments.append(
|
||||
{
|
||||
"id": len(all_segments),
|
||||
"seek": seek,
|
||||
"start": start,
|
||||
"end": end,
|
||||
"text": text,
|
||||
"tokens": text_tokens.tolist(),
|
||||
"temperature": result.temperature,
|
||||
"avg_logprob": result.avg_logprob,
|
||||
"compression_ratio": result.compression_ratio,
|
||||
"no_speech_prob": result.no_speech_prob,
|
||||
}
|
||||
)
|
||||
if verbose:
|
||||
print(f"[{format_timestamp(start)} --> {format_timestamp(end)}] {text}")
|
||||
|
||||
if carry_initial_prompt:
|
||||
nignored = max(len(initial_prompt_tokens), prompt_reset_since)
|
||||
remaining_prompt = all_tokens[nignored:][-remaining_prompt_length:]
|
||||
decode_options["prompt"] = initial_prompt_tokens + remaining_prompt
|
||||
else:
|
||||
decode_options["prompt"] = all_tokens[prompt_reset_since:]
|
||||
# show the progress bar when verbose is False (otherwise the transcribed text will be printed)
|
||||
num_frames = mel.shape[-1]
|
||||
previous_seek_value = seek
|
||||
|
||||
result: DecodingResult = decode_with_fallback(mel_segment)
|
||||
with tqdm.tqdm(total=num_frames, unit='frames', disable=verbose is not False) as pbar:
|
||||
while seek < num_frames:
|
||||
timestamp_offset = float(seek * HOP_LENGTH / SAMPLE_RATE)
|
||||
segment = pad_or_trim(mel[:, seek:], N_FRAMES).to(model.device).to(dtype)
|
||||
segment_duration = segment.shape[-1] * HOP_LENGTH / SAMPLE_RATE
|
||||
|
||||
decode_options["prompt"] = all_tokens[prompt_reset_since:]
|
||||
result: DecodingResult = decode_with_fallback(segment)
|
||||
tokens = torch.tensor(result.tokens)
|
||||
|
||||
if no_speech_threshold is not None:
|
||||
# no voice activity check
|
||||
should_skip = result.no_speech_prob > no_speech_threshold
|
||||
if (
|
||||
logprob_threshold is not None
|
||||
and result.avg_logprob > logprob_threshold
|
||||
):
|
||||
if logprob_threshold is not None and result.avg_logprob > logprob_threshold:
|
||||
# don't skip if the logprob is high enough, despite the no_speech_prob
|
||||
should_skip = False
|
||||
|
||||
if should_skip:
|
||||
seek += segment_size # fast-forward to the next segment boundary
|
||||
seek += segment.shape[-1] # fast-forward to the next segment boundary
|
||||
continue
|
||||
|
||||
previous_seek = seek
|
||||
current_segments = []
|
||||
|
||||
# anomalous words are very long/short/improbable
|
||||
def word_anomaly_score(word: dict) -> float:
|
||||
probability = word.get("probability", 0.0)
|
||||
duration = word["end"] - word["start"]
|
||||
score = 0.0
|
||||
if probability < 0.15:
|
||||
score += 1.0
|
||||
if duration < 0.133:
|
||||
score += (0.133 - duration) * 15
|
||||
if duration > 2.0:
|
||||
score += duration - 2.0
|
||||
return score
|
||||
|
||||
def is_segment_anomaly(segment: Optional[dict]) -> bool:
|
||||
if segment is None or not segment["words"]:
|
||||
return False
|
||||
words = [w for w in segment["words"] if w["word"] not in punctuation]
|
||||
words = words[:8]
|
||||
score = sum(word_anomaly_score(w) for w in words)
|
||||
return score >= 3 or score + 0.01 >= len(words)
|
||||
|
||||
def next_words_segment(segments: List[dict]) -> Optional[dict]:
|
||||
return next((s for s in segments if s["words"]), None)
|
||||
|
||||
timestamp_tokens: torch.Tensor = tokens.ge(tokenizer.timestamp_begin)
|
||||
single_timestamp_ending = timestamp_tokens[-2:].tolist() == [False, True]
|
||||
|
||||
consecutive = torch.where(timestamp_tokens[:-1] & timestamp_tokens[1:])[0]
|
||||
consecutive.add_(1)
|
||||
if len(consecutive) > 0:
|
||||
# if the output contains two consecutive timestamp tokens
|
||||
slices = consecutive.tolist()
|
||||
if single_timestamp_ending:
|
||||
slices.append(len(tokens))
|
||||
|
||||
consecutive = torch.where(timestamp_tokens[:-1] & timestamp_tokens[1:])[0].add_(1)
|
||||
if len(consecutive) > 0: # if the output contains two consecutive timestamp tokens
|
||||
last_slice = 0
|
||||
for current_slice in slices:
|
||||
for current_slice in consecutive:
|
||||
sliced_tokens = tokens[last_slice:current_slice]
|
||||
start_timestamp_pos = (
|
||||
start_timestamp_position = (
|
||||
sliced_tokens[0].item() - tokenizer.timestamp_begin
|
||||
)
|
||||
end_timestamp_pos = (
|
||||
end_timestamp_position = (
|
||||
sliced_tokens[-1].item() - tokenizer.timestamp_begin
|
||||
)
|
||||
current_segments.append(
|
||||
new_segment(
|
||||
start=time_offset + start_timestamp_pos * time_precision,
|
||||
end=time_offset + end_timestamp_pos * time_precision,
|
||||
tokens=sliced_tokens,
|
||||
result=result,
|
||||
)
|
||||
add_segment(
|
||||
start=timestamp_offset + start_timestamp_position * time_precision,
|
||||
end=timestamp_offset + end_timestamp_position * time_precision,
|
||||
text_tokens=sliced_tokens[1:-1],
|
||||
result=result,
|
||||
)
|
||||
last_slice = current_slice
|
||||
|
||||
if single_timestamp_ending:
|
||||
# single timestamp at the end means no speech after the last timestamp.
|
||||
seek += segment_size
|
||||
else:
|
||||
# otherwise, ignore the unfinished segment and seek to the last timestamp
|
||||
last_timestamp_pos = (
|
||||
tokens[last_slice - 1].item() - tokenizer.timestamp_begin
|
||||
)
|
||||
seek += last_timestamp_pos * input_stride
|
||||
last_timestamp_position = (
|
||||
tokens[last_slice - 1].item() - tokenizer.timestamp_begin
|
||||
)
|
||||
seek += last_timestamp_position * input_stride
|
||||
all_tokens.extend(tokens[: last_slice + 1].tolist())
|
||||
else:
|
||||
duration = segment_duration
|
||||
timestamps = tokens[timestamp_tokens.nonzero().flatten()]
|
||||
if (
|
||||
len(timestamps) > 0
|
||||
and timestamps[-1].item() != tokenizer.timestamp_begin
|
||||
):
|
||||
if len(timestamps) > 0 and timestamps[-1].item() != tokenizer.timestamp_begin:
|
||||
# no consecutive timestamps but it has a timestamp; use the last one.
|
||||
last_timestamp_pos = (
|
||||
timestamps[-1].item() - tokenizer.timestamp_begin
|
||||
)
|
||||
duration = last_timestamp_pos * time_precision
|
||||
# single timestamp at the end means no speech after the last timestamp.
|
||||
last_timestamp_position = timestamps[-1].item() - tokenizer.timestamp_begin
|
||||
duration = last_timestamp_position * time_precision
|
||||
|
||||
current_segments.append(
|
||||
new_segment(
|
||||
start=time_offset,
|
||||
end=time_offset + duration,
|
||||
tokens=tokens,
|
||||
result=result,
|
||||
)
|
||||
)
|
||||
seek += segment_size
|
||||
|
||||
if word_timestamps:
|
||||
add_word_timestamps(
|
||||
segments=current_segments,
|
||||
model=model,
|
||||
tokenizer=tokenizer,
|
||||
mel=mel_segment,
|
||||
num_frames=segment_size,
|
||||
prepend_punctuations=prepend_punctuations,
|
||||
append_punctuations=append_punctuations,
|
||||
last_speech_timestamp=last_speech_timestamp,
|
||||
add_segment(
|
||||
start=timestamp_offset,
|
||||
end=timestamp_offset + duration,
|
||||
text_tokens=tokens,
|
||||
result=result,
|
||||
)
|
||||
|
||||
if not single_timestamp_ending:
|
||||
last_word_end = get_end(current_segments)
|
||||
if last_word_end is not None and last_word_end > time_offset:
|
||||
seek = round(last_word_end * FRAMES_PER_SECOND)
|
||||
|
||||
# skip silence before possible hallucinations
|
||||
if hallucination_silence_threshold is not None:
|
||||
threshold = hallucination_silence_threshold
|
||||
if not single_timestamp_ending:
|
||||
last_word_end = get_end(current_segments)
|
||||
if last_word_end is not None and last_word_end > time_offset:
|
||||
remaining_duration = window_end_time - last_word_end
|
||||
if remaining_duration > threshold:
|
||||
seek = round(last_word_end * FRAMES_PER_SECOND)
|
||||
else:
|
||||
seek = previous_seek + segment_size
|
||||
|
||||
# if first segment might be a hallucination, skip leading silence
|
||||
first_segment = next_words_segment(current_segments)
|
||||
if first_segment is not None and is_segment_anomaly(first_segment):
|
||||
gap = first_segment["start"] - time_offset
|
||||
if gap > threshold:
|
||||
seek = previous_seek + round(gap * FRAMES_PER_SECOND)
|
||||
continue
|
||||
|
||||
# skip silence before any possible hallucination that is surrounded
|
||||
# by silence or more hallucinations
|
||||
hal_last_end = last_speech_timestamp
|
||||
for si in range(len(current_segments)):
|
||||
segment = current_segments[si]
|
||||
if not segment["words"]:
|
||||
continue
|
||||
if is_segment_anomaly(segment):
|
||||
next_segment = next_words_segment(
|
||||
current_segments[si + 1 :]
|
||||
)
|
||||
if next_segment is not None:
|
||||
hal_next_start = next_segment["words"][0]["start"]
|
||||
else:
|
||||
hal_next_start = time_offset + segment_duration
|
||||
silence_before = (
|
||||
segment["start"] - hal_last_end > threshold
|
||||
or segment["start"] < threshold
|
||||
or segment["start"] - time_offset < 2.0
|
||||
)
|
||||
silence_after = (
|
||||
hal_next_start - segment["end"] > threshold
|
||||
or is_segment_anomaly(next_segment)
|
||||
or window_end_time - segment["end"] < 2.0
|
||||
)
|
||||
if silence_before and silence_after:
|
||||
seek = round(
|
||||
max(time_offset + 1, segment["start"])
|
||||
* FRAMES_PER_SECOND
|
||||
)
|
||||
if content_duration - segment["end"] < threshold:
|
||||
seek = content_frames
|
||||
current_segments[si:] = []
|
||||
break
|
||||
hal_last_end = segment["end"]
|
||||
|
||||
last_word_end = get_end(current_segments)
|
||||
if last_word_end is not None:
|
||||
last_speech_timestamp = last_word_end
|
||||
|
||||
if verbose:
|
||||
for segment in current_segments:
|
||||
start, end, text = segment["start"], segment["end"], segment["text"]
|
||||
line = f"[{format_timestamp(start)} --> {format_timestamp(end)}] {text}"
|
||||
print(make_safe(line))
|
||||
|
||||
# if a segment is instantaneous or does not contain text, clear it
|
||||
for i, segment in enumerate(current_segments):
|
||||
if segment["start"] == segment["end"] or segment["text"].strip() == "":
|
||||
segment["text"] = ""
|
||||
segment["tokens"] = []
|
||||
segment["words"] = []
|
||||
|
||||
all_segments.extend(
|
||||
[
|
||||
{"id": i, **segment}
|
||||
for i, segment in enumerate(
|
||||
current_segments, start=len(all_segments)
|
||||
)
|
||||
]
|
||||
)
|
||||
all_tokens.extend(
|
||||
[token for segment in current_segments for token in segment["tokens"]]
|
||||
)
|
||||
seek += segment.shape[-1]
|
||||
all_tokens.extend(tokens.tolist())
|
||||
|
||||
if not condition_on_previous_text or result.temperature > 0.5:
|
||||
# do not feed the prompt tokens if a high temperature was used
|
||||
prompt_reset_since = len(all_tokens)
|
||||
|
||||
# update progress bar
|
||||
pbar.update(min(content_frames, seek) - previous_seek)
|
||||
pbar.update(min(num_frames, seek) - previous_seek_value)
|
||||
previous_seek_value = seek
|
||||
|
||||
return dict(
|
||||
text=tokenizer.decode(all_tokens[len(initial_prompt_tokens) :]),
|
||||
segments=all_segments,
|
||||
language=language,
|
||||
)
|
||||
return dict(text=tokenizer.decode(all_tokens[len(initial_prompt):]), segments=all_segments, language=language)
|
||||
|
||||
|
||||
def cli():
|
||||
from . import available_models
|
||||
|
||||
def valid_model_name(name):
|
||||
if name in available_models() or os.path.exists(name):
|
||||
return name
|
||||
raise ValueError(
|
||||
f"model should be one of {available_models()} or path to a model checkpoint"
|
||||
)
|
||||
|
||||
# fmt: off
|
||||
parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
|
||||
parser.add_argument("audio", nargs="+", type=str, help="audio file(s) to transcribe")
|
||||
parser.add_argument("--model", default="turbo", type=valid_model_name, help="name of the Whisper model to use")
|
||||
parser.add_argument("--model", default="small", choices=available_models(), help="name of the Whisper model to use")
|
||||
parser.add_argument("--model_dir", type=str, default=None, help="the path to save model files; uses ~/.cache/whisper by default")
|
||||
parser.add_argument("--device", default="cuda" if torch.cuda.is_available() else "cpu", help="device to use for PyTorch inference")
|
||||
parser.add_argument("--output_dir", "-o", type=str, default=".", help="directory to save the outputs")
|
||||
parser.add_argument("--output_format", "-f", type=str, default="all", choices=["txt", "vtt", "srt", "tsv", "json", "all"], help="format of the output file; if not specified, all available formats will be produced")
|
||||
parser.add_argument("--verbose", type=str2bool, default=True, help="whether to print out the progress and debug messages")
|
||||
|
||||
parser.add_argument("--task", type=str, default="transcribe", choices=["transcribe", "translate"], help="whether to perform X->X speech recognition ('transcribe') or X->English translation ('translate')")
|
||||
@ -545,8 +268,6 @@ def cli():
|
||||
|
||||
parser.add_argument("--suppress_tokens", type=str, default="-1", help="comma-separated list of token ids to suppress during sampling; '-1' will suppress most special characters except common punctuations")
|
||||
parser.add_argument("--initial_prompt", type=str, default=None, help="optional text to provide as a prompt for the first window.")
|
||||
parser.add_argument("--carry_initial_prompt", type=str2bool, default=False, help="if True, prepend initial_prompt to every internal decode() call. May reduce the effectiveness of condition_on_previous_text")
|
||||
|
||||
parser.add_argument("--condition_on_previous_text", type=str2bool, default=True, help="if True, provide the previous output of the model as a prompt for the next window; disabling may make the text inconsistent across windows, but the model becomes less prone to getting stuck in a failure loop")
|
||||
parser.add_argument("--fp16", type=str2bool, default=True, help="whether to perform inference in fp16; True by default")
|
||||
|
||||
@ -554,70 +275,51 @@ def cli():
|
||||
parser.add_argument("--compression_ratio_threshold", type=optional_float, default=2.4, help="if the gzip compression ratio is higher than this value, treat the decoding as failed")
|
||||
parser.add_argument("--logprob_threshold", type=optional_float, default=-1.0, help="if the average log probability is lower than this value, treat the decoding as failed")
|
||||
parser.add_argument("--no_speech_threshold", type=optional_float, default=0.6, help="if the probability of the <|nospeech|> token is higher than this value AND the decoding has failed due to `logprob_threshold`, consider the segment as silence")
|
||||
parser.add_argument("--word_timestamps", type=str2bool, default=False, help="(experimental) extract word-level timestamps and refine the results based on them")
|
||||
parser.add_argument("--prepend_punctuations", type=str, default="\"\'“¿([{-", help="if word_timestamps is True, merge these punctuation symbols with the next word")
|
||||
parser.add_argument("--append_punctuations", type=str, default="\"\'.。,，!！?？:：”)]}、", help="if word_timestamps is True, merge these punctuation symbols with the previous word")
|
||||
parser.add_argument("--highlight_words", type=str2bool, default=False, help="(requires --word_timestamps True) underline each word as it is spoken in srt and vtt")
|
||||
parser.add_argument("--max_line_width", type=optional_int, default=None, help="(requires --word_timestamps True) the maximum number of characters in a line before breaking the line")
|
||||
parser.add_argument("--max_line_count", type=optional_int, default=None, help="(requires --word_timestamps True) the maximum number of lines in a segment")
|
||||
parser.add_argument("--max_words_per_line", type=optional_int, default=None, help="(requires --word_timestamps True, no effect with --max_line_width) the maximum number of words in a segment")
|
||||
parser.add_argument("--threads", type=optional_int, default=0, help="number of threads used by torch for CPU inference; supercedes MKL_NUM_THREADS/OMP_NUM_THREADS")
|
||||
parser.add_argument("--clip_timestamps", type=str, default="0", help="comma-separated list start,end,start,end,... timestamps (in seconds) of clips to process, where the last end timestamp defaults to the end of the file")
|
||||
parser.add_argument("--hallucination_silence_threshold", type=optional_float, help="(requires --word_timestamps True) skip silent periods longer than this threshold (in seconds) when a possible hallucination is detected")
|
||||
# fmt: on
|
||||
|
||||
args = parser.parse_args().__dict__
|
||||
model_name: str = args.pop("model")
|
||||
model_dir: str = args.pop("model_dir")
|
||||
output_dir: str = args.pop("output_dir")
|
||||
output_format: str = args.pop("output_format")
|
||||
device: str = args.pop("device")
|
||||
os.makedirs(output_dir, exist_ok=True)
|
||||
|
||||
if model_name.endswith(".en") and args["language"] not in {"en", "English"}:
|
||||
if args["language"] is not None:
|
||||
warnings.warn(
|
||||
f"{model_name} is an English-only model but receipted '{args['language']}'; using English instead."
|
||||
)
|
||||
warnings.warn(f"{model_name} is an English-only model but receipted '{args['language']}'; using English instead.")
|
||||
args["language"] = "en"
|
||||
|
||||
temperature = args.pop("temperature")
|
||||
if (increment := args.pop("temperature_increment_on_fallback")) is not None:
|
||||
temperature = tuple(np.arange(temperature, 1.0 + 1e-6, increment))
|
||||
temperature_increment_on_fallback = args.pop("temperature_increment_on_fallback")
|
||||
if temperature_increment_on_fallback is not None:
|
||||
temperature = tuple(np.arange(temperature, 1.0 + 1e-6, temperature_increment_on_fallback))
|
||||
else:
|
||||
temperature = [temperature]
|
||||
|
||||
if (threads := args.pop("threads")) > 0:
|
||||
threads = args.pop("threads")
|
||||
if threads > 0:
|
||||
torch.set_num_threads(threads)
|
||||
|
||||
from . import load_model
|
||||
|
||||
model = load_model(model_name, device=device, download_root=model_dir)
|
||||
|
||||
writer = get_writer(output_format, output_dir)
|
||||
word_options = [
|
||||
"highlight_words",
|
||||
"max_line_count",
|
||||
"max_line_width",
|
||||
"max_words_per_line",
|
||||
]
|
||||
if not args["word_timestamps"]:
|
||||
for option in word_options:
|
||||
if args[option]:
|
||||
parser.error(f"--{option} requires --word_timestamps True")
|
||||
if args["max_line_count"] and not args["max_line_width"]:
|
||||
warnings.warn("--max_line_count has no effect without --max_line_width")
|
||||
if args["max_words_per_line"] and args["max_line_width"]:
|
||||
warnings.warn("--max_words_per_line has no effect with --max_line_width")
|
||||
writer_args = {arg: args.pop(arg) for arg in word_options}
|
||||
for audio_path in args.pop("audio"):
|
||||
try:
|
||||
result = transcribe(model, audio_path, temperature=temperature, **args)
|
||||
writer(result, audio_path, **writer_args)
|
||||
except Exception as e:
|
||||
traceback.print_exc()
|
||||
print(f"Skipping {audio_path} due to {type(e).__name__}: {str(e)}")
|
||||
result = transcribe(model, audio_path, temperature=temperature, **args)
|
||||
|
||||
audio_basename = os.path.basename(audio_path)
|
||||
|
||||
# save TXT
|
||||
with open(os.path.join(output_dir, audio_basename + ".txt"), "w", encoding="utf-8") as txt:
|
||||
write_txt(result["segments"], file=txt)
|
||||
|
||||
# save VTT
|
||||
with open(os.path.join(output_dir, audio_basename + ".vtt"), "w", encoding="utf-8") as vtt:
|
||||
write_vtt(result["segments"], file=vtt)
|
||||
|
||||
# save SRT
|
||||
with open(os.path.join(output_dir, audio_basename + ".srt"), "w", encoding="utf-8") as srt:
|
||||
write_srt(result["segments"], file=srt)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
if __name__ == '__main__':
|
||||
cli()
|
||||
|
||||
@ -1,117 +0,0 @@
|
||||
from functools import lru_cache
|
||||
|
||||
import numpy as np
|
||||
import torch
|
||||
|
||||
try:
|
||||
import triton
|
||||
import triton.language as tl
|
||||
except ImportError:
|
||||
raise RuntimeError("triton import failed; try `pip install --pre triton`")
|
||||
|
||||
|
||||
@triton.jit
|
||||
def dtw_kernel(
|
||||
cost, trace, x, x_stride, cost_stride, trace_stride, N, M, BLOCK_SIZE: tl.constexpr
|
||||
):
|
||||
offsets = tl.arange(0, BLOCK_SIZE)
|
||||
mask = offsets < M
|
||||
|
||||
for k in range(1, N + M + 1): # k = i + j
|
||||
tl.debug_barrier()
|
||||
|
||||
p0 = cost + (k - 1) * cost_stride
|
||||
p1 = cost + k * cost_stride
|
||||
p2 = cost + k * cost_stride + 1
|
||||
|
||||
c0 = tl.load(p0 + offsets, mask=mask)
|
||||
c1 = tl.load(p1 + offsets, mask=mask)
|
||||
c2 = tl.load(p2 + offsets, mask=mask)
|
||||
|
||||
x_row = tl.load(x + (k - 1) * x_stride + offsets, mask=mask, other=0)
|
||||
cost_row = x_row + tl.minimum(tl.minimum(c0, c1), c2)
|
||||
|
||||
cost_ptr = cost + (k + 1) * cost_stride + 1
|
||||
tl.store(cost_ptr + offsets, cost_row, mask=mask)
|
||||
|
||||
trace_ptr = trace + (k + 1) * trace_stride + 1
|
||||
tl.store(trace_ptr + offsets, 2, mask=mask & (c2 <= c0) & (c2 <= c1))
|
||||
tl.store(trace_ptr + offsets, 1, mask=mask & (c1 <= c0) & (c1 <= c2))
|
||||
tl.store(trace_ptr + offsets, 0, mask=mask & (c0 <= c1) & (c0 <= c2))
|
||||
|
||||
|
||||
@lru_cache(maxsize=None)
|
||||
def median_kernel(filter_width: int):
|
||||
@triton.jit
|
||||
def kernel(
|
||||
y, x, x_stride, y_stride, BLOCK_SIZE: tl.constexpr
|
||||
): # x.shape[-1] == filter_width
|
||||
row_idx = tl.program_id(0)
|
||||
offsets = tl.arange(0, BLOCK_SIZE)
|
||||
mask = offsets < y_stride
|
||||
|
||||
x_ptr = x + row_idx * x_stride # noqa: F841
|
||||
y_ptr = y + row_idx * y_stride
|
||||
|
||||
LOAD_ALL_ROWS_HERE # noqa: F821
|
||||
|
||||
BUBBLESORT_HERE # noqa: F821
|
||||
|
||||
tl.store(y_ptr + offsets, MIDDLE_ROW_HERE, mask=mask) # noqa: F821
|
||||
|
||||
kernel = triton.JITFunction(kernel.fn)
|
||||
new_kernel = kernel.src.replace(
|
||||
" LOAD_ALL_ROWS_HERE",
|
||||
"\n".join(
|
||||
[
|
||||
f" row{i} = tl.load(x_ptr + offsets + {i}, mask=mask)"
|
||||
for i in range(filter_width)
|
||||
]
|
||||
),
|
||||
)
|
||||
|
||||
new_kernel = new_kernel.replace(
|
||||
" BUBBLESORT_HERE",
|
||||
"\n\n".join(
|
||||
[
|
||||
"\n\n".join(
|
||||
[
|
||||
"\n".join(
|
||||
[
|
||||
f" smaller = tl.where(row{j} < row{j + 1}, row{j}, row{j + 1})",
|
||||
f" larger = tl.where(row{j} > row{j + 1}, row{j}, row{j + 1})",
|
||||
f" row{j} = smaller",
|
||||
f" row{j + 1} = larger",
|
||||
]
|
||||
)
|
||||
for j in range(filter_width - i - 1)
|
||||
]
|
||||
)
|
||||
for i in range(filter_width // 2 + 1)
|
||||
]
|
||||
),
|
||||
)
|
||||
|
||||
new_kernel = new_kernel.replace("MIDDLE_ROW_HERE", f"row{filter_width // 2}")
|
||||
|
||||
if hasattr(kernel, "_unsafe_update_src") is True:
|
||||
kernel._unsafe_update_src(new_kernel)
|
||||
kernel.hash = None
|
||||
else:
|
||||
kernel.src = new_kernel
|
||||
|
||||
return kernel
|
||||
|
||||
|
||||
def median_filter_cuda(x: torch.Tensor, filter_width: int):
|
||||
"""Apply a median filter of given width along the last dimension of x"""
|
||||
slices = x.contiguous().unfold(-1, filter_width, 1)
|
||||
grid = np.prod(slices.shape[:-2])
|
||||
|
||||
kernel = median_kernel(filter_width)
|
||||
y = torch.empty_like(slices[..., 0])
|
||||
|
||||
BLOCK_SIZE = 1 << (y.stride(-2) - 1).bit_length()
|
||||
kernel[(grid,)](y, x, x.stride(-2), y.stride(-2), BLOCK_SIZE=BLOCK_SIZE)
|
||||
|
||||
return y
|
||||
302
whisper/utils.py
302
whisper/utils.py
@ -1,24 +1,5 @@
|
||||
import json
|
||||
import os
|
||||
import re
|
||||
import sys
|
||||
import zlib
|
||||
from typing import Callable, List, Optional, TextIO
|
||||
|
||||
system_encoding = sys.getdefaultencoding()
|
||||
|
||||
if system_encoding != "utf-8":
|
||||
|
||||
def make_safe(string):
|
||||
# replaces any character not representable using the system default encoding with an '?',
|
||||
# avoiding UnicodeEncodeError (https://github.com/openai/whisper/discussions/729).
|
||||
return string.encode(system_encoding, errors="replace").decode(system_encoding)
|
||||
|
||||
else:
|
||||
|
||||
def make_safe(string):
|
||||
# utf-8 can encode any Unicode code point, so no need to do the round-trip encoding
|
||||
return string
|
||||
from typing import Iterator, TextIO
|
||||
|
||||
|
||||
def exact_div(x, y):
|
||||
@ -47,9 +28,7 @@ def compression_ratio(text) -> float:
|
||||
return len(text_bytes) / len(zlib.compress(text_bytes))
|
||||
|
||||
|
||||
def format_timestamp(
|
||||
seconds: float, always_include_hours: bool = False, decimal_marker: str = "."
|
||||
):
|
||||
def format_timestamp(seconds: float, always_include_hours: bool = False, decimal_marker: str = '.'):
|
||||
assert seconds >= 0, "non-negative timestamp expected"
|
||||
milliseconds = round(seconds * 1000.0)
|
||||
|
||||
@ -63,256 +42,47 @@ def format_timestamp(
|
||||
milliseconds -= seconds * 1_000
|
||||
|
||||
hours_marker = f"{hours:02d}:" if always_include_hours or hours > 0 else ""
|
||||
return (
|
||||
f"{hours_marker}{minutes:02d}:{seconds:02d}{decimal_marker}{milliseconds:03d}"
|
||||
)
|
||||
return f"{hours_marker}{minutes:02d}:{seconds:02d}{decimal_marker}{milliseconds:03d}"
|
||||
|
||||
|
||||
def get_start(segments: List[dict]) -> Optional[float]:
|
||||
return next(
|
||||
(w["start"] for s in segments for w in s["words"]),
|
||||
segments[0]["start"] if segments else None,
|
||||
)
|
||||
def write_txt(transcript: Iterator[dict], file: TextIO):
|
||||
for segment in transcript:
|
||||
print(segment['text'].strip(), file=file, flush=True)
|
||||
|
||||
|
||||
def get_end(segments: List[dict]) -> Optional[float]:
|
||||
return next(
|
||||
(w["end"] for s in reversed(segments) for w in reversed(s["words"])),
|
||||
segments[-1]["end"] if segments else None,
|
||||
)
|
||||
|
||||
|
||||
class ResultWriter:
|
||||
extension: str
|
||||
|
||||
def __init__(self, output_dir: str):
|
||||
self.output_dir = output_dir
|
||||
|
||||
def __call__(
|
||||
self, result: dict, audio_path: str, options: Optional[dict] = None, **kwargs
|
||||
):
|
||||
audio_basename = os.path.basename(audio_path)
|
||||
audio_basename = os.path.splitext(audio_basename)[0]
|
||||
output_path = os.path.join(
|
||||
self.output_dir, audio_basename + "." + self.extension
|
||||
)
|
||||
|
||||
with open(output_path, "w", encoding="utf-8") as f:
|
||||
self.write_result(result, file=f, options=options, **kwargs)
|
||||
|
||||
def write_result(
|
||||
self, result: dict, file: TextIO, options: Optional[dict] = None, **kwargs
|
||||
):
|
||||
raise NotImplementedError
|
||||
|
||||
|
||||
class WriteTXT(ResultWriter):
|
||||
extension: str = "txt"
|
||||
|
||||
def write_result(
|
||||
self, result: dict, file: TextIO, options: Optional[dict] = None, **kwargs
|
||||
):
|
||||
for segment in result["segments"]:
|
||||
print(segment["text"].strip(), file=file, flush=True)
|
||||
|
||||
|
||||
class SubtitlesWriter(ResultWriter):
|
||||
always_include_hours: bool
|
||||
decimal_marker: str
|
||||
|
||||
def iterate_result(
|
||||
self,
|
||||
result: dict,
|
||||
options: Optional[dict] = None,
|
||||
*,
|
||||
max_line_width: Optional[int] = None,
|
||||
max_line_count: Optional[int] = None,
|
||||
highlight_words: bool = False,
|
||||
max_words_per_line: Optional[int] = None,
|
||||
):
|
||||
options = options or {}
|
||||
max_line_width = max_line_width or options.get("max_line_width")
|
||||
max_line_count = max_line_count or options.get("max_line_count")
|
||||
highlight_words = highlight_words or options.get("highlight_words", False)
|
||||
max_words_per_line = max_words_per_line or options.get("max_words_per_line")
|
||||
preserve_segments = max_line_count is None or max_line_width is None
|
||||
max_line_width = max_line_width or 1000
|
||||
max_words_per_line = max_words_per_line or 1000
|
||||
|
||||
def iterate_subtitles():
|
||||
line_len = 0
|
||||
line_count = 1
|
||||
# the next subtitle to yield (a list of word timings with whitespace)
|
||||
subtitle: List[dict] = []
|
||||
last: float = get_start(result["segments"]) or 0.0
|
||||
for segment in result["segments"]:
|
||||
chunk_index = 0
|
||||
words_count = max_words_per_line
|
||||
while chunk_index < len(segment["words"]):
|
||||
remaining_words = len(segment["words"]) - chunk_index
|
||||
if max_words_per_line > len(segment["words"]) - chunk_index:
|
||||
words_count = remaining_words
|
||||
for i, original_timing in enumerate(
|
||||
segment["words"][chunk_index : chunk_index + words_count]
|
||||
):
|
||||
timing = original_timing.copy()
|
||||
long_pause = (
|
||||
not preserve_segments and timing["start"] - last > 3.0
|
||||
)
|
||||
has_room = line_len + len(timing["word"]) <= max_line_width
|
||||
seg_break = i == 0 and len(subtitle) > 0 and preserve_segments
|
||||
if (
|
||||
line_len > 0
|
||||
and has_room
|
||||
and not long_pause
|
||||
and not seg_break
|
||||
):
|
||||
# line continuation
|
||||
line_len += len(timing["word"])
|
||||
else:
|
||||
# new line
|
||||
timing["word"] = timing["word"].strip()
|
||||
if (
|
||||
len(subtitle) > 0
|
||||
and max_line_count is not None
|
||||
and (long_pause or line_count >= max_line_count)
|
||||
or seg_break
|
||||
):
|
||||
# subtitle break
|
||||
yield subtitle
|
||||
subtitle = []
|
||||
line_count = 1
|
||||
elif line_len > 0:
|
||||
# line break
|
||||
line_count += 1
|
||||
timing["word"] = "\n" + timing["word"]
|
||||
line_len = len(timing["word"].strip())
|
||||
subtitle.append(timing)
|
||||
last = timing["start"]
|
||||
chunk_index += max_words_per_line
|
||||
if len(subtitle) > 0:
|
||||
yield subtitle
|
||||
|
||||
if len(result["segments"]) > 0 and "words" in result["segments"][0]:
|
||||
for subtitle in iterate_subtitles():
|
||||
subtitle_start = self.format_timestamp(subtitle[0]["start"])
|
||||
subtitle_end = self.format_timestamp(subtitle[-1]["end"])
|
||||
subtitle_text = "".join([word["word"] for word in subtitle])
|
||||
if highlight_words:
|
||||
last = subtitle_start
|
||||
all_words = [timing["word"] for timing in subtitle]
|
||||
for i, this_word in enumerate(subtitle):
|
||||
start = self.format_timestamp(this_word["start"])
|
||||
end = self.format_timestamp(this_word["end"])
|
||||
if last != start:
|
||||
yield last, start, subtitle_text
|
||||
|
||||
yield start, end, "".join(
|
||||
[
|
||||
(
|
||||
re.sub(r"^(\s*)(.*)$", r"\1<u>\2</u>", word)
|
||||
if j == i
|
||||
else word
|
||||
)
|
||||
for j, word in enumerate(all_words)
|
||||
]
|
||||
)
|
||||
last = end
|
||||
else:
|
||||
yield subtitle_start, subtitle_end, subtitle_text
|
||||
else:
|
||||
for segment in result["segments"]:
|
||||
segment_start = self.format_timestamp(segment["start"])
|
||||
segment_end = self.format_timestamp(segment["end"])
|
||||
segment_text = segment["text"].strip().replace("-->", "->")
|
||||
yield segment_start, segment_end, segment_text
|
||||
|
||||
def format_timestamp(self, seconds: float):
|
||||
return format_timestamp(
|
||||
seconds=seconds,
|
||||
always_include_hours=self.always_include_hours,
|
||||
decimal_marker=self.decimal_marker,
|
||||
def write_vtt(transcript: Iterator[dict], file: TextIO):
|
||||
print("WEBVTT\n", file=file)
|
||||
for segment in transcript:
|
||||
print(
|
||||
f"{format_timestamp(segment['start'])} --> {format_timestamp(segment['end'])}\n"
|
||||
f"{segment['text'].strip().replace('-->', '->')}\n",
|
||||
file=file,
|
||||
flush=True,
|
||||
)
|
||||
|
||||
|
||||
class WriteVTT(SubtitlesWriter):
|
||||
extension: str = "vtt"
|
||||
always_include_hours: bool = False
|
||||
decimal_marker: str = "."
|
||||
|
||||
def write_result(
|
||||
self, result: dict, file: TextIO, options: Optional[dict] = None, **kwargs
|
||||
):
|
||||
print("WEBVTT\n", file=file)
|
||||
for start, end, text in self.iterate_result(result, options, **kwargs):
|
||||
print(f"{start} --> {end}\n{text}\n", file=file, flush=True)
|
||||
|
||||
|
||||
class WriteSRT(SubtitlesWriter):
|
||||
extension: str = "srt"
|
||||
always_include_hours: bool = True
|
||||
decimal_marker: str = ","
|
||||
|
||||
def write_result(
|
||||
self, result: dict, file: TextIO, options: Optional[dict] = None, **kwargs
|
||||
):
|
||||
for i, (start, end, text) in enumerate(
|
||||
self.iterate_result(result, options, **kwargs), start=1
|
||||
):
|
||||
print(f"{i}\n{start} --> {end}\n{text}\n", file=file, flush=True)
|
||||
|
||||
|
||||
class WriteTSV(ResultWriter):
|
||||
def write_srt(transcript: Iterator[dict], file: TextIO):
|
||||
"""
|
||||
Write a transcript to a file in TSV (tab-separated values) format containing lines like:
|
||||
<start time in integer milliseconds>\t<end time in integer milliseconds>\t<transcript text>
|
||||
Write a transcript to a file in SRT format.
|
||||
|
||||
Using integer milliseconds as start and end times means there's no chance of interference from
|
||||
an environment setting a language encoding that causes the decimal in a floating point number
|
||||
to appear as a comma; also is faster and more efficient to parse & store, e.g., in C++.
|
||||
Example usage:
|
||||
from pathlib import Path
|
||||
from whisper.utils import write_srt
|
||||
|
||||
result = transcribe(model, audio_path, temperature=temperature, **args)
|
||||
|
||||
# save SRT
|
||||
audio_basename = Path(audio_path).stem
|
||||
with open(Path(output_dir) / (audio_basename + ".srt"), "w", encoding="utf-8") as srt:
|
||||
write_srt(result["segments"], file=srt)
|
||||
"""
|
||||
|
||||
extension: str = "tsv"
|
||||
|
||||
def write_result(
|
||||
self, result: dict, file: TextIO, options: Optional[dict] = None, **kwargs
|
||||
):
|
||||
print("start", "end", "text", sep="\t", file=file)
|
||||
for segment in result["segments"]:
|
||||
print(round(1000 * segment["start"]), file=file, end="\t")
|
||||
print(round(1000 * segment["end"]), file=file, end="\t")
|
||||
print(segment["text"].strip().replace("\t", " "), file=file, flush=True)
|
||||
|
||||
|
||||
class WriteJSON(ResultWriter):
|
||||
extension: str = "json"
|
||||
|
||||
def write_result(
|
||||
self, result: dict, file: TextIO, options: Optional[dict] = None, **kwargs
|
||||
):
|
||||
json.dump(result, file)
|
||||
|
||||
|
||||
def get_writer(
|
||||
output_format: str, output_dir: str
|
||||
) -> Callable[[dict, TextIO, dict], None]:
|
||||
writers = {
|
||||
"txt": WriteTXT,
|
||||
"vtt": WriteVTT,
|
||||
"srt": WriteSRT,
|
||||
"tsv": WriteTSV,
|
||||
"json": WriteJSON,
|
||||
}
|
||||
|
||||
if output_format == "all":
|
||||
all_writers = [writer(output_dir) for writer in writers.values()]
|
||||
|
||||
def write_all(
|
||||
result: dict, file: TextIO, options: Optional[dict] = None, **kwargs
|
||||
):
|
||||
for writer in all_writers:
|
||||
writer(result, file, options, **kwargs)
|
||||
|
||||
return write_all
|
||||
|
||||
return writers[output_format](output_dir)
|
||||
for i, segment in enumerate(transcript, start=1):
|
||||
# write srt lines
|
||||
print(
|
||||
f"{i}\n"
|
||||
f"{format_timestamp(segment['start'], always_include_hours=True, decimal_marker=',')} --> "
|
||||
f"{format_timestamp(segment['end'], always_include_hours=True, decimal_marker=',')}\n"
|
||||
f"{segment['text'].strip().replace('-->', '->')}\n",
|
||||
file=file,
|
||||
flush=True,
|
||||
)
|
||||
|
||||
@ -1 +1 @@
|
||||
__version__ = "20250625"
|
||||
__version__ = "20230117"
|
||||
|
||||
Loading…
x
Reference in New Issue
Block a user