Repository navigation
CI Report:
https://ci-tests.linuxserver.io/linuxserver/faster-whisper/v3.9.1-ls16/index.html
LinuxServer Changes:
Full Changelog: v3.9.0-ls15...v3.9.1-ls16
Remote Changes:
-
A custom
--modeladvertises all of Whisper's languages again, as it did in
3.5.0. Since 3.9.0 it was given the language list of its backend's default
model, which says nothing about a model the server cannot inspect without
downloading it: a customonnx-asrmodel was advertised as Russian only
(GigaAM's list), and customsherpaandfunasrmodels were narrowed the
same way. Home Assistant then would not offer the model to a pipeline in its
own language, for example a Korean onnx-asr model to a Korean pipeline
(#148). Default models and.encheckpoints still advertise exactly what they
support -
The
onnx-asrandqwen3-asrbackends no longer keep CPU cores busy after a
transcript is sent. onnxruntime's thread pools spin while waiting for more
work by default, which slows down whatever starts on the same cores next,
such as an LLM answering the command (#135). With onnx-asr's Parakeet on a
desktop CPU, each transcript left about half a core-second of spinning in the
second after it; that is now zero. onnx-asr transcribes no slower, and
qwen3-asr, which runs its decoder once per token, about 2% slower- The
sherpabackend still spins: sherpa-onnx creates its onnxruntime
sessions itself and does not expose the setting
- The
-
The
funasrextra now installs kaldi-native-fbank. FunASR computes
SenseVoice's features with torchaudio or, in recent releases,
kaldi-native-fbank, but declares neither, so an install with torch and
without torchaudio (for example alongside thetransformersextra) failed
every SenseVoice transcription withImportError. torch itself still has to
be installed separately. The Home Assistant app and the Docker image were not
affected