Merge branch 'main' into main

.github/workflows/python-publish.yml

@@ -17,11 +17,11 @@ jobs:
     - name: Set up Python
       uses: actions/setup-python@v5
       with:
-        python-version: '3.8'
+        python-version: '3.12'
     - name: Install dependencies
       run: |
         python -m pip install --upgrade pip
-        pip install setuptools wheel twine
+        pip install setuptools wheel twine build
     - name: Release
       if: ${{ steps.regex-match.outputs.match != '' }}
       uses: softprops/action-gh-release@v2

CHANGELOG.md

@@ -1,5 +1,26 @@
 # 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))

README.md

@@ -77,25 +77,35 @@ Whisper's performance varies widely depending on the language. The figure below
 
 ![WER breakdown by language](https://github.com/openai/whisper/assets/266841/f4619d66-1058-4005-8f67-a9d811b77c62)
 
-
-
 ## Command-line usage
 
 The following command will transcribe speech in audio files, using the `turbo` model:
 
-    whisper audio.flac audio.mp3 audio.wav --model turbo
+```bash
+whisper audio.flac audio.mp3 audio.wav --model turbo
+```
 
-The default setting (which selects the `turbo` model) works well for transcribing English. To transcribe an audio file containing non-English speech, you can specify the language using the `--language` option:
+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`. 
 
-    whisper japanese.wav --language Japanese
+For example, to transcribe an audio file containing non-English speech, you can specify the language:
 
-Adding `--task translate` will translate the speech into English:
+```bash
+whisper japanese.wav --language Japanese
+```
 
-    whisper japanese.wav --language Japanese --task translate
+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.
 
 Run the following to view all available options:
 
-    whisper --help
+```bash
+whisper --help
+```
 
 See [tokenizer.py](https://github.com/openai/whisper/blob/main/whisper/tokenizer.py) for the list of all available languages.
 

notebooks/LibriSpeech.ipynb

@@ -949,7 +949,8 @@
       "style": "IPY_MODEL_039b53f2702c4179af7e0548018d0588",
       "value": " 164/164 [05:08<00:00,  1.86s/it]"
      }
-    }
+    },
+    "state": {}
    }
   }
  },

notebooks/Multilingual_ASR.ipynb

@@ -4219,7 +4219,8 @@
             "_view_name": "StyleView",
             "description_width": ""
           }
-        }
+        },
+        "state": {}
       }
     }
   },

whisper/__init__.py

@@ -147,7 +147,8 @@ def load_model(
     with (
         io.BytesIO(checkpoint_file) if in_memory else open(checkpoint_file, "rb")
     ) as fp:
-        checkpoint = torch.load(fp, map_location=device)
+        kwargs = {"weights_only": True} if torch.__version__ >= "1.13" else {}
+        checkpoint = torch.load(fp, map_location=device, **kwargs)
     del checkpoint_file
 
     dims = ModelDimensions(**checkpoint["dims"])

whisper/timing.py

@@ -117,7 +117,7 @@ def dtw_cuda(x, BLOCK_SIZE=1024):
     x_skew = x_skew.T.contiguous()
     cost = torch.ones(N + M + 2, M + 2) * np.inf
     cost[0, 0] = 0
-    cost = cost.cuda()
+    cost = cost.to(x.device)
     trace = torch.zeros_like(cost, dtype=torch.int32)
 
     dtw_kernel[(1,)](

whisper/triton_ops.py

@@ -60,7 +60,7 @@ def median_kernel(filter_width: int):
         tl.store(y_ptr + offsets, MIDDLE_ROW_HERE, mask=mask)  # noqa: F821
 
     kernel = triton.JITFunction(kernel.fn)
-    kernel.src = kernel.src.replace(
+    new_kernel = kernel.src.replace(
         "    LOAD_ALL_ROWS_HERE",
         "\n".join(
             [
@@ -69,7 +69,8 @@ def median_kernel(filter_width: int):
             ]
         ),
     )
-    kernel.src = kernel.src.replace(
+
+    new_kernel = new_kernel.replace(
         "    BUBBLESORT_HERE",
         "\n\n".join(
             [
@@ -90,7 +91,14 @@ def median_kernel(filter_width: int):
             ]
         ),
     )
-    kernel.src = kernel.src.replace("MIDDLE_ROW_HERE", f"row{filter_width // 2}")
+
+    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
 

whisper/version.py

@@ -1 +1 @@
-__version__ = "20240930"
+__version__ = "20250625"