sci-ml/pyannote-audio
State-of-the-art speaker diarization toolkit (PyTorch)
Screenshots
View
Download
BrowseChangeLog
commit 4893a2b1e02d359cb23708f37d2ffb0ad7a6e3c4
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Wed Jul 1 09:15:08 2026 +0200
sci-ml/pyannote-audio: drop 4.0.4
commit f8d3ffc3671b67b0f8a1a0d659547d098e5bab83
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Wed Jul 1 09:15:07 2026 +0200
sci-ml/pyannote-audio: add 4.0.7
commit adfa03a37cb22c5ad93e890557d1b47877a9eea7
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Tue Jun 23 00:52:39 2026 +0200
sci-ml/pyannote-audio: add 4.0.5
commit 4bdef7bb4c1a996a2184fd7a2a1e45c6e7582f89
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Thu Jun 18 18:51:43 2026 +0200
sci-ml/pyannote-audio: Keyword 4.0.4 for ~arm64
Speaker-diarization toolkit on PyTorch; the torch backend (caffe2/
pytorch 2.11) runs on aarch64 and pyannote model architectures
instantiate with live torch params. Pure-Python; aarch64-portable.
commit 09a3601e682611210790bfead087f5016a7ba07e
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Mon May 18 19:43:44 2026 +0200
sci-ml/pyannote-audio: use canonical PyPI slug (pyannote-audio)
pypi.org/project/pyannote.audio/ permanently redirects to
/pyannote-audio/. Switch HOMEPAGE + metadata.xml pypi remote-id
to the hyphen form. PYPI_PN stays "pyannote.audio" (the sdist
filename still uses the dot form).
commit 4138521faf073de9deffcaa3e515c1acdf265ae3
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sat May 16 10:52:40 2026 +0200
sci-ml/pyannote-audio: new package, add 4.0.4
State-of-the-art speaker diarization, voice activity detection,
overlapped speech detection, and speaker embedding toolkit.
Sits on top of the freshly-landed lightning/torchcodec/torchmetrics/
pyannote-{core,database,metrics,pipeline}/pyannoteai-sdk stack
(16 sibling deps in this commit series).
The Python package alone is enough to import; running the
pretrained pyannote/speaker-diarization-3.1 pipeline at runtime
additionally requires a HuggingFace token and accepting the model
terms on huggingface.co/pyannote — that's a per-user setup step,
not something the ebuild can do.
SINGLE_IMPL via pytorch. Tests restricted (papermill + downloaded
HF model bundle); local build-check verified with USE=-cuda.
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Wed Jul 1 09:15:08 2026 +0200
sci-ml/pyannote-audio: drop 4.0.4
commit f8d3ffc3671b67b0f8a1a0d659547d098e5bab83
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Wed Jul 1 09:15:07 2026 +0200
sci-ml/pyannote-audio: add 4.0.7
commit adfa03a37cb22c5ad93e890557d1b47877a9eea7
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Tue Jun 23 00:52:39 2026 +0200
sci-ml/pyannote-audio: add 4.0.5
commit 4bdef7bb4c1a996a2184fd7a2a1e45c6e7582f89
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Thu Jun 18 18:51:43 2026 +0200
sci-ml/pyannote-audio: Keyword 4.0.4 for ~arm64
Speaker-diarization toolkit on PyTorch; the torch backend (caffe2/
pytorch 2.11) runs on aarch64 and pyannote model architectures
instantiate with live torch params. Pure-Python; aarch64-portable.
commit 09a3601e682611210790bfead087f5016a7ba07e
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Mon May 18 19:43:44 2026 +0200
sci-ml/pyannote-audio: use canonical PyPI slug (pyannote-audio)
pypi.org/project/pyannote.audio/ permanently redirects to
/pyannote-audio/. Switch HOMEPAGE + metadata.xml pypi remote-id
to the hyphen form. PYPI_PN stays "pyannote.audio" (the sdist
filename still uses the dot form).
commit 4138521faf073de9deffcaa3e515c1acdf265ae3
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sat May 16 10:52:40 2026 +0200
sci-ml/pyannote-audio: new package, add 4.0.4
State-of-the-art speaker diarization, voice activity detection,
overlapped speech detection, and speaker embedding toolkit.
Sits on top of the freshly-landed lightning/torchcodec/torchmetrics/
pyannote-{core,database,metrics,pipeline}/pyannoteai-sdk stack
(16 sibling deps in this commit series).
The Python package alone is enough to import; running the
pretrained pyannote/speaker-diarization-3.1 pipeline at runtime
additionally requires a HuggingFace token and accepting the model
terms on huggingface.co/pyannote — that's a per-user setup step,
not something the ebuild can do.
SINGLE_IMPL via pytorch. Tests restricted (papermill + downloaded
HF model bundle); local build-check verified with USE=-cuda.