VAD

Voice activity detection backed by Silero VAD (ONNX).

from voiceclean.vad import VAD

Requires: pip install onnxruntime

`VAD`

class VAD(
    sample_rate: int = 8000,
    threshold: float = 0.5,
)

Silero VAD is an ONNX model (~2 MB) that runs in <1 ms per chunk on CPU. Supports 8 kHz and 16 kHz natively.

The model is downloaded automatically on first use to ~/.cache/voiceclean/silero_vad.onnx.

Parameters:

Parameter	Type	Default	Description
`sample_rate`	`int`	`8000`	Audio sample rate. Must be 8000 or 16000.
`threshold`	`float`	`0.5`	Speech probability threshold

Methods

`process`

def process(self, audio: bytes) -> VADResult

Detect speech in audio. Buffers audio internally and returns the VAD result for the most recent complete frame.

Frame sizes:

8000 Hz: 256 samples (32 ms)
16000 Hz: 512 samples (32 ms)

Parameters:

audio — Raw PCM int16 mono bytes at the configured sample rate.

Returns: VADResult

`reset`

def reset(self) -> None

Reset internal LSTM state and audio buffer. Call between conversations to avoid state leaking across calls.

`VADResult`

@dataclass
class VADResult:
    is_speech: bool
    speech_prob: float

Field	Type	Description
`is_speech`	`bool`	`True` if `speech_prob >= threshold`
`speech_prob`	`float`	Speech probability from 0.0 to 1.0

Example

from voiceclean.vad import VAD

vad = VAD(sample_rate=8000, threshold=0.5)

result = vad.process(audio_chunk)
if result.is_speech:
    print(f"Speech detected (confidence: {result.speech_prob:.2f})")

# Between calls, reset state
vad.reset()

Model details

Property	Value
Model	Silero VAD v5
Format	ONNX
Size	~2 MB
Inference time	<1 ms per frame (CPU)
Internal state	LSTM, shape (2, 1, 128)
Source	snakers4/silero-vad
License	MIT