sci-ml/sherpa-onnx-bin
Speech-to-text, TTS, speaker diarization etc. using onnxruntime (binary wheels)
Screenshots
View
Download
BrowseChangeLog
commit 94bfc9a219cb882fbd7e508fce2fb790e63df3ab
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sat May 16 13:41:35 2026 +0200
sci-ml/sherpa-onnx-bin: new package, add 1.13.2
ONNX-Runtime-based speech stack from k2-fsa: ASR, TTS, speaker
diarization, VAD, source separation, KWS. Standalone alternative
to sci-ml/pyannote-audio for users who don't want the pytorch
stack — runs CPU-only, uses ONNX-converted pyannote+3D-Speaker
models that aren't HuggingFace-gated.
-bin route via upstream manylinux2014 wheels. Two co-installable
wheels: sherpa-onnx-core (Python-version-agnostic C++ libs) +
sherpa-onnx (per-CPython-ABI bindings); SRC_URI gates the latter
on PYTHON_SINGLE_TARGET. Both unpack into the same sherpa_onnx/
tree so the binding's RPATH=$ORIGIN dlopen resolves libonnxruntime.so.
From-source path would require packaging dev-libs/onnxruntime first
(absent from ::gentoo, multi-hundred-MB C++ build with abseil/
protobuf/etc.) — defer until there's a concrete need.
Build + import smoke-tested on python3_13; OfflineSpeakerDiarization
class loads end-to-end. Runtime model files (pyannote-segmentation-3.0
ONNX + 3D-Speaker embeddings) live on HuggingFace, ungated.
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sat May 16 13:41:35 2026 +0200
sci-ml/sherpa-onnx-bin: new package, add 1.13.2
ONNX-Runtime-based speech stack from k2-fsa: ASR, TTS, speaker
diarization, VAD, source separation, KWS. Standalone alternative
to sci-ml/pyannote-audio for users who don't want the pytorch
stack — runs CPU-only, uses ONNX-converted pyannote+3D-Speaker
models that aren't HuggingFace-gated.
-bin route via upstream manylinux2014 wheels. Two co-installable
wheels: sherpa-onnx-core (Python-version-agnostic C++ libs) +
sherpa-onnx (per-CPython-ABI bindings); SRC_URI gates the latter
on PYTHON_SINGLE_TARGET. Both unpack into the same sherpa_onnx/
tree so the binding's RPATH=$ORIGIN dlopen resolves libonnxruntime.so.
From-source path would require packaging dev-libs/onnxruntime first
(absent from ::gentoo, multi-hundred-MB C++ build with abseil/
protobuf/etc.) — defer until there's a concrete need.
Build + import smoke-tested on python3_13; OfflineSpeakerDiarization
class loads end-to-end. Runtime model files (pyannote-segmentation-3.0
ONNX + 3D-Speaker embeddings) live on HuggingFace, ungated.