Merge branch 'main' into main

.github/workflows/test.yml

@@ -41,16 +41,29 @@ jobs:
     runs-on: ubuntu-latest
     strategy:
       matrix:
-        python-version: ['3.8', '3.9', '3.10', '3.11']
+        include:
-        pytorch-version: [1.13.1, 2.0.0]
+          - python-version: '3.8'
-        exclude:
-          - python-version: '3.11'
             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: '3.10'
+            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'"
     steps:
       - uses: conda-incubator/setup-miniconda@v2
       - run: conda install -n test ffmpeg python=${{ matrix.python-version }}
-      - run: pip3 install torch==${{ matrix.pytorch-version }}+cpu --index-url https://download.pytorch.org/whl/cpu
       - uses: actions/checkout@v3
       - run: echo "$CONDA/envs/test/bin" >> $GITHUB_PATH
-      - run: pip install .["dev"]
+      - 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'

CHANGELOG.md

@@ -1,5 +1,19 @@
 # CHANGELOG
 
+## [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))

README.md

@@ -57,17 +57,21 @@ pip install setuptools-rust
 
 ## Available models and languages
 
-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 inference speed relative to the large model; actual speed may vary depending on many factors including the available hardware.
+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.
 
 |  Size  | Parameters | English-only model | Multilingual model | Required VRAM | Relative speed |
 |:------:|:----------:|:------------------:|:------------------:|:-------------:|:--------------:|
-|  tiny  |    39 M    |     `tiny.en`      |       `tiny`       |     ~1 GB     |      ~32x      |
+|  tiny  |    39 M    |     `tiny.en`      |       `tiny`       |     ~1 GB     |      ~10x      |
-|  base  |    74 M    |     `base.en`      |       `base`       |     ~1 GB     |      ~16x      |
+|  base  |    74 M    |     `base.en`      |       `base`       |     ~1 GB     |      ~7x       |
-| small  |   244 M    |     `small.en`     |      `small`       |     ~2 GB     |      ~6x       |
+| small  |   244 M    |     `small.en`     |      `small`       |     ~2 GB     |      ~4x       |
 | 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.
 
 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.
 
@@ -77,9 +81,9 @@ Whisper's performance varies widely depending on the language. The figure below
 
 ## Command-line usage
 
-The following command will transcribe speech in audio files, using the `medium` model:
+The following command will transcribe speech in audio files, using the `turbo` model:
 
-    whisper audio.flac audio.mp3 audio.wav --model medium
+    whisper audio.flac audio.mp3 audio.wav --model 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:
 
@@ -103,7 +107,7 @@ Transcription can also be performed within Python:
 ```python
 import whisper
 
-model = whisper.load_model("base")
+model = whisper.load_model("turbo")
 result = model.transcribe("audio.mp3")
 print(result["text"])
 ```
@@ -115,7 +119,7 @@ Below is an example usage of `whisper.detect_language()` and `whisper.decode()`
 ```python
 import whisper
 
-model = whisper.load_model("base")
+model = whisper.load_model("turbo")
 
 # load audio and pad/trim it to fit 30 seconds
 audio = whisper.load_audio("audio.mp3")

model-card.md

@@ -16,13 +16,15 @@ 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.
 
 
 ### Release date
 
-September 2022 (original series), December 2022 (`large-v2`), and November 2023 (`large-v3`)
+September 2022 (original series), December 2022 (`large-v2`), November 2023 (`large-v3`), September 2024 (`large-v3-turbo`)
 
 ### Model type
 

requirements.txt

@@ -4,4 +4,4 @@ torch
 tqdm
 more-itertools
 tiktoken
-triton>=2.0.0,<3;platform_machine=="x86_64" and sys_platform=="linux" or sys_platform=="linux2"
+triton>=2.0.0;platform_machine=="x86_64" and sys_platform=="linux" or sys_platform=="linux2"

setup.py

@@ -13,7 +13,7 @@ def read_version(fname="whisper/version.py"):
 
 requirements = []
 if sys.platform.startswith("linux") and platform.machine() == "x86_64":
-    requirements.append("triton>=2.0.0,<3")
+    requirements.append("triton>=2.0.0")
 
 setup(
     name="openai-whisper",

whisper/__init__.py

@@ -27,6 +27,8 @@ _MODELS = {
     "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
@@ -44,6 +46,8 @@ _ALIGNMENT_HEADS = {
     "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`",
 }
 
 

whisper/model.py

@@ -1,7 +1,8 @@
 import base64
 import gzip
+from contextlib import contextmanager
 from dataclasses import dataclass
-from typing import Dict, Iterable, Optional
+from typing import Dict, Iterable, Optional, Tuple
 
 import numpy as np
 import torch
@@ -12,6 +13,14 @@ 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
+
 
 @dataclass
 class ModelDimensions:
@@ -59,7 +68,19 @@ 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
@@ -92,20 +113,30 @@ class MultiHeadAttention(nn.Module):
 
     def qkv_attention(
         self, q: Tensor, k: Tensor, v: Tensor, mask: Optional[Tensor] = None
-    ):
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
         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) * scale
+        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, 3, 1) * scale
+        k = k.view(*k.shape[:2], self.n_head, -1).permute(0, 2, 1, 3)
         v = v.view(*v.shape[:2], self.n_head, -1).permute(0, 2, 1, 3)
 
-        qk = q @ k
+        if SDPA_AVAILABLE and MultiHeadAttention.use_sdpa:
-        if mask is not None:
+            a = scaled_dot_product_attention(
-            qk = qk + mask[:n_ctx, :n_ctx]
+                q, k, v, is_causal=mask is not None and n_ctx > 1
-        qk = qk.float()
+            )
+            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()
 
-        w = F.softmax(qk, dim=-1).to(q.dtype)
+            w = F.softmax(qk, dim=-1).to(q.dtype)
-        return (w @ v).permute(0, 2, 1, 3).flatten(start_dim=2), qk.detach()
+            out = (w @ v).permute(0, 2, 1, 3).flatten(start_dim=2)
+            qk = qk.detach()
+
+        return out, qk
 
 
 class ResidualAttentionBlock(nn.Module):

whisper/timing.py

@@ -191,7 +191,9 @@ def find_alignment(
         for i, block in enumerate(model.decoder.blocks)
     ]
 
-    with torch.no_grad():
+    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)

whisper/transcribe.py

@@ -874,7 +874,7 @@ def cli():
     # 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="small", type=valid_model_name, help="name of the Whisper model to use")
+    parser.add_argument("--model", default="turbo", type=valid_model_name, 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")

whisper/version.py

@@ -1 +1 @@
-__version__ = "20231117"
+__version__ = "20240930"