Release 20250625

* Fix: Update torch.load to use weights_only=True to prevent security warning

* Update __init__.py

* Update __init__.py

---------

Co-authored-by: Jong Wook Kim <jongwook@openai.com>

.github/dependabot.yml

@@ -0,0 +1,13 @@
+# 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

.github/workflows/python-publish.yml

@@ -8,23 +8,23 @@ jobs:
   deploy:
     runs-on: ubuntu-latest
     steps:
-    - uses: actions/checkout@v3
+    - uses: actions/checkout@v4
     - 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@v4
+      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: ${{ steps.regex-match.outputs.match != '' }}
-      uses: softprops/action-gh-release@v1
+      uses: softprops/action-gh-release@v2
       with:
         tag_name: v${{ 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 setup.py sdist
+        python -m build --sdist
         twine upload dist/*

.github/workflows/test.yml

@@ -11,19 +11,19 @@ jobs:
   pre-commit:
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v3
+      - uses: actions/checkout@v4
       - name: Fetch base branch
         run: git fetch origin ${{ github.base_ref }}
-      - uses: actions/setup-python@v4
+      - uses: actions/setup-python@v5
         with:
-          python-version: "3.8"
+          python-version: "3.9"
           architecture: x64
       - name: Get pip cache dir
         id: pip-cache
         run: |
           echo "dir=$(pip cache dir)" >> $GITHUB_OUTPUT
       - name: pip/pre-commit cache
-        uses: actions/cache@v3
+        uses: actions/cache@v4
         with:
           path: |
             ${{ steps.pip-cache.outputs.dir }}
@@ -33,15 +33,19 @@ jobs:
             ${{ runner.os }}-pip-pre-commit
       - name: pre-commit
         run: |
-          pip install -U pre-commit
+          pip install --upgrade pre-commit
           pre-commit install --install-hooks
           pre-commit run --all-files
   whisper-test:
     needs: pre-commit
     runs-on: ubuntu-latest
     strategy:
+      fail-fast: false
       matrix:
         include:
+          - python-version: '3.8'
+            pytorch-version: 1.10.1
+            numpy-requirement: "'numpy<2'"
           - python-version: '3.8'
             pytorch-version: 1.13.1
             numpy-requirement: "'numpy<2'"
@@ -60,10 +64,16 @@ jobs:
           - python-version: '3.12'
             pytorch-version: 2.4.1
             numpy-requirement: "'numpy'"
+          - python-version: '3.12'
+            pytorch-version: 2.5.1
+            numpy-requirement: "'numpy'"
+          - python-version: '3.13'
+            pytorch-version: 2.5.1
+            numpy-requirement: "'numpy'"
     steps:
-      - uses: conda-incubator/setup-miniconda@v2
+      - uses: conda-incubator/setup-miniconda@v3
       - run: conda install -n test ffmpeg python=${{ matrix.python-version }}
-      - uses: actions/checkout@v3
+      - uses: actions/checkout@v4
       - 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'

.pre-commit-config.yaml

@@ -1,6 +1,6 @@
 repos:
   - repo: https://github.com/pre-commit/pre-commit-hooks
-    rev: v4.0.1
+    rev: v5.0.0
     hooks:
       - id: check-json
       - id: end-of-file-fixer
@@ -11,17 +11,17 @@ repos:
       - id: check-added-large-files
         args: [--maxkb=4096]
   - repo: https://github.com/psf/black
-    rev: 23.7.0
+    rev: 25.1.0
     hooks:
       - id: black
   - repo: https://github.com/pycqa/isort
-    rev: 5.12.0
+    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: 6.0.0
+    rev: 7.1.1
     hooks:
       - id: flake8
         types: [python]

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 `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:
+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.
 
@@ -126,7 +136,7 @@ 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).to(model.device)
+mel = whisper.log_mel_spectrogram(audio, n_mels=model.dims.n_mels).to(model.device)
 
 # detect the spoken language
 _, probs = model.detect_language(mel)

data/README.md

@@ -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 ad 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 and 2 of the [s5b recipe](https://github.com/kaldi-asr/kaldi/tree/master/egs/ami/s5b).
 
 
 ## Long-form English-only datasets

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": {}
       }
     }
   },

pyproject.toml

@@ -1,3 +1,50 @@
+[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]
@@ -5,4 +52,3 @@ profile = "black"
 include_trailing_comma = true
 line_length = 88
 multi_line_output = 3
-

setup.py

@@ -1,42 +0,0 @@
-import platform
-import sys
-from pathlib import Path
-
-import pkg_resources
-from setuptools import find_packages, setup
-
-
-def read_version(fname="whisper/version.py"):
-    exec(compile(open(fname, encoding="utf-8").read(), fname, "exec"))
-    return locals()["__version__"]
-
-
-requirements = []
-if sys.platform.startswith("linux") and platform.machine() == "x86_64":
-    requirements.append("triton>=2.0.0")
-
-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=">=3.8",
-    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(
-            Path(__file__).with_name("requirements.txt").open()
-        )
-    ],
-    entry_points={
-        "console_scripts": ["whisper=whisper.transcribe:cli"],
-    },
-    include_package_data=True,
-    extras_require={"dev": ["pytest", "scipy", "black", "flake8", "isort"]},
-)

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/audio.py

@@ -122,7 +122,7 @@ 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 is supported
+        The number of Mel-frequency filters, only 80 and 128 are supported
 
     padding: int
         Number of zero samples to pad to the right
@@ -132,7 +132,7 @@ def log_mel_spectrogram(
 
     Returns
     -------
-    torch.Tensor, shape = (80, n_frames)
+    torch.Tensor, shape = (n_mels, n_frames)
         A Tensor that contains the Mel spectrogram
     """
     if not torch.is_tensor(audio):

whisper/normalizers/basic.py

@@ -30,15 +30,19 @@ 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]
+            else (
+                ADDITIONAL_DIACRITICS[c]
                 if c in ADDITIONAL_DIACRITICS
-        else ""
+                else (
+                    ""
                     if unicodedata.category(c) == "Mn"
-        else " "
+                    else " " if unicodedata.category(c)[0] in "MSP" else c
-        if unicodedata.category(c)[0] in "MSP"
+                )
-        else c
+            )
+        )
         for c in unicodedata.normalize("NFKD", s)
     )
 

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/transcribe.py

@@ -46,6 +46,7 @@ def transcribe(
     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 = "\"'.。,，!！?？:：”)]}、",
@@ -102,6 +103,11 @@ def transcribe(
         "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
 
@@ -208,6 +214,8 @@ def transcribe(
             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:
@@ -227,9 +235,11 @@ 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 = []
 
@@ -275,7 +285,13 @@ def transcribe(
             segment_duration = segment_size * HOP_LENGTH / SAMPLE_RATE
             mel_segment = pad_or_trim(mel_segment, N_FRAMES).to(model.device).to(dtype)
 
+            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:]
+
             result: DecodingResult = decode_with_fallback(mel_segment)
             tokens = torch.tensor(result.tokens)
 
@@ -529,6 +545,8 @@ 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")
 

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/utils.py

@@ -209,9 +209,11 @@ class SubtitlesWriter(ResultWriter):
 
                         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)
                             ]
                         )

whisper/version.py

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