Detect Voice Activity (Silero VAD)

How it fits

Typical backends

Caveats

• I/O contractOutput spans are in seconds relative to the start of the input AudioFrame, not absolute stream time, so carry a sibling timestamp or accumulate a frame offset downstream if absolute times are needed. Multi-channel input is mixed to mono before resampling.
• I/O contractEvery input AudioFrame emits exactly one span list and never drops a message. An empty frame returns an empty list without invoking the model; silence returns an empty list; an unexpected fault (such as a non-positive sample rate) is logged and also returns an empty list.
• AccuracySilero is trained at 16 kHz and the component resamples internally; supplying audio captured below 16 kHz hurts accuracy noticeably.
• Parameter interactionthreshold is the per-frame speech-probability cutoff, min_speech_duration_ms drops too-short spans, min_silence_duration_ms sets how much silence splits one utterance into two, and speech_pad_ms pads the final spans. All four are re-read on every tick, so live tuning takes effect on the next call without a restart.
• Fallbackdevice is captured once at startup and switching CPU/GPU needs a redeploy; a value starting with `cuda` silently falls back to CPU when CUDA is unavailable.