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. Carry a sibling timestamp or accumulate a frame offset downstream if absolute times are needed.
• Hard constraintAn empty input AudioFrame returns an empty list without invoking the model. A non-positive sample rate aborts mid-tick with an error.
• I/O contractMulti-channel input is mixed to mono before resampling.
• AccuracySilero is trained at 16 kHz. The component resamples internally; downsampling from below 16 kHz hurts accuracy noticeably.
• Parameter interactionthreshold, min_speech_duration_ms, min_silence_duration_ms, and speech_pad_ms are re-read on every tick, so live tuning takes effect on the next call without restarting.
• Parameter interactionthreshold controls per-frame speech probability cutoff. min_speech_duration_ms drops too-short spans after merging. min_silence_duration_ms controls how much silence is required to split an utterance into two spans. speech_pad_ms pads the final spans after all other processing.
• Fallbackdevice is captured once at startup; switching CPU/GPU requires a redeploy. device starting with `cuda` silently falls back to CPU with a warning when CUDA is unavailable.