ragleap-voice v0.3.6 release notes

ragleap-voice 0.3.6 (pre-alpha). Every model is still unverified and none is a default.

Added

ragleap_voice.vad_silero.SileroGate: an optional model-based voice-activity detector with the same feed() and finish() interface and events as StreamGate. It needs the new extra vad (pip install "ragleap-voice[vad]", which adds onnxruntime and numpy; the core stays standard library only).
It is not exported, not used by default and not wired into transcribe(gate=...) or transcribe_stream. The model file is not bundled or downloaded: you supply the path of silero_vad_16k_op15.onnx, and a file without the pinned SHA-256 is refused. The model has no entry in models.toml, and its weights licence and training data are not verified.

Documentation and benchmarks (each one run, limits stated in benchmarks/voice/README.md)

Per-call timing of transcribe() and Transcriber on a Tesla T4, and a check that the GPU was used.
Word error on clean speech against ffmpeg mu-law and ragleap's own 8 kHz mu-law and A-law path (English, Spanish, Russian, 30 clips each): no difference between ragleap's path and ffmpeg's was found at that sample size.
False accepts on real recorded noise (MUSAN, free-sound folder, no speech mixed in): StreamGate about 99%, Silero 4.4% to 5.4%. Nothing about speech detection is claimed.
docs/verification-criteria.md and a smoke test record for SileroGate. With the default settings it splits speech at short pauses; a better value has not been measured.