Detect Voice Activity (Silero VAD)

Frame-level voice activity detection using the Silero VAD model. Returns true whenever any speech region is detected inside the audio frame — typical use is gating ASR or downstream language stages.

How it fits

AudioFrame  ──►  detect_voice_activity_silero_vad  ──► Bool (speech present)

Pick this when you want a cheap "is anyone talking?" signal to avoid spending compute on silence. For full turn-end detection use detect_turn_speech.

Typical pipelines

• Speech-gated transcription: input_audio_file → denoise_audio_mp_senet → detect_voice_activity_silero_vad → filter → transcribe_audio_faster_whisper → send_http.
• Conversation segment collector: live audio → detect_voice_activity_silero_vad → collect (gate) → transcribe_audio_faster_whisper → generate_text_ollama.
• Recording auto-trim: input_audio_file → detect_voice_activity_silero_vad → filter → output_audio_file (only speech).

Caveats

• The boolean fires on ANY speech-flagged region in the frame, no matter how short — clicks, coughs, or sneezes can register as speech.
• Silero is trained at 16 kHz. The worker resamples internally; that's fine for higher rates but downsampling from 8 kHz hurts accuracy.
• No exposed tuning knobs (threshold, min/max speech duration, padding). If you need to filter by duration, accumulate VAD outputs downstream.
• On CPU the worker disables PyTorch's NNPACK backend (Silero needs that for predictable behaviour). On GPU this is irrelevant.
• Only device is configurable in this worker — for fine-grained Silero tuning you'll need a custom variant or post-processing.

Related components

• detect_turn_speech — paired component for "speaker has just finished" detection.
• denoise_audio_mp_senet — typical pre-VAD step on noisy inputs.
• transcribe_audio_faster_whisper, transcribe_audio_moonshine, transcribe_audio_parakeet, transcribe_audio_sensevoice — typical downstream ASR.
• filter, collect — typical control-flow consumers.