fasr-vad-silero 0.5.3

Silero VAD model for fasr

fasr-vad-silero

Chinese documentation

Streaming Silero VAD for fasr. The model is loaded through torch.hub and wrapped as a fasr streaming VADModel that emits AudioChunk objects with segment_start, segment_mid, and segment_end states.

Install

pip install fasr-vad-silero

Registered Model

Registry name	Class	Best for
stream_silero	SileroStreamVAD	Lightweight streaming VAD

Streaming Usage

from fasr.config import registry

vad = registry.vad_models.get("stream_silero")(
    threshold=0.5,
    silence_duration_ms=400,
)

for input_chunk in audio_chunks:
    for speech_chunk in vad.push_chunk(input_chunk):
        print(speech_chunk.vad_state, speech_chunk.start_ms, speech_chunk.end_ms)

Quick choices:

Goal	Use	Result
Reduce noise triggers	threshold=0.65	Requires higher speech probability
Keep quiet speech	threshold=0.35	More sensitive, with more false-positive risk
End speech sooner	silence_duration_ms=200	Lower endpoint latency
Avoid chopping pauses	silence_duration_ms=700	More tolerant of short pauses

Confection Config

[vad_model]
@vad_models = "stream_silero"
threshold = 0.5
silence_duration_ms = 400
prefix_padding_ms = 0
sample_rate = 16000
chunk_size_ms = 32

Parameters

Parameter	Type / range	Default	Higher value	Lower value	Change when
threshold	float, 0.0 to 1.0	0.5	More conservative; fewer noise starts	More sensitive; more weak speech	Noise triggers starts, or quiet speech is missed
silence_duration_ms	int >= 0	400	Longer pauses before ending speech	Faster endpoint	Speech is chopped, or endpoint is late
prefix_padding_ms	int >= 0	0	Keeps more leading context and reduces prefix clipping	Keeps segments tighter but raises the risk of clipping the first syllable	The beginning of an utterance is cut off, or too much pre-roll noise is included
sample_rate	int	16000	Keep at 16 kHz for Silero	Keep at 16 kHz for Silero	Usually do not change
chunk_size_ms	int	32	Larger input chunks, fewer calls	Smaller chunks, lower latency	Realtime scheduling needs tuning

Tuning Guide

Symptom	Try first
Background noise starts speech	Raise threshold to 0.6 or 0.7
Quiet speech is missed	Lower threshold to 0.35 or 0.4
Speech ends too late	Lower silence_duration_ms to 200 or 300
Speech is split during pauses	Raise silence_duration_ms to 600 or 800
The first syllable is clipped	Raise prefix_padding_ms to 64 or 128

Runtime Turn Detection

stream_silero supports realtime turn-detection updates through apply_turn_detection(...). This hook is intended for websocket-style server sessions where VAD behavior may be adjusted after the session has started.

Runtime-adjustable fields:

• silence_duration_ms
• threshold
• prefix_padding_ms

prefix_padding_ms is implemented in the wrapper state machine: when speech is first detected, the emitted segment start is moved earlier by the configured amount while clamping to the beginning of the buffered audio.

Dependencies

• fasr
• numpy >= 1.24
• torch >= 2.0.0
• Python 3.10-3.12