Changelog

Complete version history of Whisper releases, including new features, bug fixes, and improvements.

​

v20250625

​

Security & Bug Fixes

• Update torch.load to use weights_only=True to prevent security vulnerabilities (#2451)
• Ensure DTW cost tensor is on the same device as input tensor (#2561)
• Fixed triton kernel update to support latest triton versions (#2588)
• Fix GitHub display errors for Jupyter notebooks (#2589)
• Bugfix: Illogical “Avoid computing higher temperatures on no_speech” (#1903)
• Fix typo in data/README.md (#2433)

​

Documentation

• Updated README to specify translation model limitation (#2547)
• Updating README and doc strings to reflect that n_mels can now be 128 (#2049)
• Update README.md (#2379)

​

Features

• Add option to carry initial_prompt with the sliding window (#2343)

​

Development & CI

• Bump the github-actions group with 3 updates (#2592)
• Keep GitHub Actions up to date with GitHub’s Dependabot (#2486)
• pre-commit: Upgrade black v25.1.0 and isort v6.0.0 (#2514)
• GitHub Actions: Add Python 3.13 to the testing (#2487)
• PEP 621: Migrate from setup.py to pyproject.toml (#2435)
• pre-commit autoupdate && pre-commit run —all-files (#2484)
• Upgrade GitHub Actions (#2430)
• More pytorch versions in tests (#2408)

---

​

v20240930

​

Features

• large-v3-turbo model - New optimized model for faster inference (#2361)
• Using SDPA (Scaled Dot Product Attention) if available (#2359)

​

Development

• Allowing numpy 2 in tests (#2362)
• Test on Python/PyTorch versions up to 3.12 and 2.4.1 (#2360)

---

​

v20240927

​

Bug Fixes

• Skip silence around hallucinations (#1838)
• Fix triton env marker (#1887)

​

Development

• Pinning numpy<2 in tests (#2332)
• Relax triton requirements for compatibility with PyTorch 2.4 and newer (#2307)

---

​

v20231117

​

Compatibility

• Relax triton requirements for compatibility with PyTorch 2.1 and newer (#1802)

---

​

v20231106

​

Features

• large-v3 - New improved large model (#1761)

---

​

v20231105

​

Features

• Add new option to generate subtitles by a specific number of words (#1729)

​

Bug Fixes

• Fix exception when an audio file with no speech is provided (#1396)
• Handling transcribe exceptions (#1682)

​

Security

• allow_pickle=False while loading mel matrix in audio.py (#1511)

​

Development

• Remove tiktoken pin (#1759)

​

Documentation

• Disambiguation of the term “relative speed” in the README (#1751)

---

​

v20230918

​

Features

• Word timing tweaks (#1559)
• Implement max line width and max line count, and make word highlighting optional (#1184)

​

Bug Fixes

• Fix condition_on_previous_text (#1224)
• Fix numba deprecation notice (#1233)
• Drop ffmpeg-python dependency and call ffmpeg directly (#1242)

​

Performance

• Avoid rearranging all caches (#1483)
• Avoid computing higher temperatures on no_speech segments (#1279)

​

Improvements

• Improve timestamp heuristics (#1461)
• Squash long words at window and sentence boundaries (#1114)

​

Development

• Add .pre-commit-config.yaml (#1528)
• Python 3.11 support (#1171)
• Dropped unused execute bit from mel_filters.npz (#1254)

​

Documentation

• Fix doc of TextDecoder (#1526)
• Update model-card.md (#1643)
• Updated README.md to provide more insight on BLEU and specific appendices (#1236)
• Update README.md to reference tiktoken (#1105)
• Update decoding.py (#1219, #1155)
• Update tokenizer.py (#1163)

---

​

v20230314

​

Features

• Use tiktoken for fast tokenization (#1044)

​

Bug Fixes

• Abort find_alignment on empty input (#1090)
• Fix truncated words list when the replacement character is decoded (#1089)
• Fix alignment between the segments and the list of words (#1087)
• Fix github language stats getting dominated by jupyter notebook (#1076)

---

​

v20230308

​

Features

• kwargs in decode() for convenience (#1061)

​

Bug Fixes

• Fix all_tokens handling that caused more repetitions and discrepancy in JSON (#1060)
• Fix typo in CHANGELOG.md

---

​

v20230307

​

Bug Fixes

• Fix the repetition/hallucination issue identified in #1046 (#1052)

​

Development

• Use triton==2.0.0 (#1053)
• Install triton in x86_64 linux only (#1051)
• Update setup.py to specify python >= 3.8 requirement

---

​

v20230306

​

Features

• Word-level timestamps in transcribe() (#869)
• Decoding improvements (#1033)

​

Bug Fixes

• Fix infinite loop caused by incorrect timestamp tokens prediction (#914)

​

Development

• Remove auxiliary audio extension (#1021)
• Apply formatting with black, isort, and flake8 (#1038)
• Drop Python 3.7 support (#889)

​

Documentation

• Update README.md (#894)

---

​

v20230124

​

Features

• Add TSV formatted output in transcript, using integer start/end time in milliseconds (#228)
• Added --output_format option (#333)

​

Bug Fixes

• Fix bug where mm is mistakenly replaced with hmm in e.g. 20mm (#659)
• Handle XDG_CACHE_HOME properly for download_root (#864)

​

Improvements

• Handle printing even if sys.stdout.buffer is not available (#887)
• Use stdout for printing transcription progress (#867)
• Print ’?’ if a letter can’t be encoded using the system default encoding (#859)

---

​

v20230117

The first versioned release available on PyPI