Silero - Zero Runtime

Silero is a voice activity detection (VAD) plugin. It tells the pipeline when the caller is speaking rather than silent, so the agent reacts to speech instead of background noise.

Usage

Import the plugin and pass it to the pipeline’s vad slot.

from zeroruntime.plugins import SileroVAD

vad = SileroVAD(threshold=0.4)

# Pipeline(vad=vad, ...)

import { SileroVAD } from '@zeroruntime/js-sdk/plugins';

const vad = SileroVAD({ threshold: 0.4 });

// Pipeline({ vad, ... })

Parameters

Constructor parameters for SileroVAD. The Python and Node JS SDKs share these field names.

Parameter	Type	Default	Description
`threshold`	`float`	`0.4`	Speech-probability cutoff for classifying a frame as speech (0-1). Lower catches softer speech but allows more false positives.
`start_threshold`	`float`	`None`	Probability needed to start a speech segment. When set, overrides `threshold`.
`end_threshold`	`float`	`None`	Probability below which a speech segment ends. When set, overrides `stop_threshold`.
`stop_threshold`	`float`	`0.25`	Probability below which a speech segment ends (used when `end_threshold` is unset).
`min_speech_duration`	`float`	`0.3`	Minimum seconds of speech before it counts as a segment.
`min_silence_duration`	`float`	`0.4`	Silence (seconds) required to mark the end of speech.
`padding_duration`	`float`	`0.5`	Audio (seconds) kept before and after each segment.
`max_buffered_speech`	`float`	`60.0`	Maximum seconds of speech buffered in memory.
`sample_rate`	`int`	`16000`	Sample rate the model runs at (used when `model_sample_rate` is unset).
`input_sample_rate`	`int`	`48000`	Sample rate (Hz) of the incoming audio.
`model_sample_rate`	`int`	`None`	Sample rate the model runs at; falls back to `sample_rate` when unset.
`smoothing_factor`	`float`	`0.35`	EMA smoothing weight applied to speech probabilities.
`force_cpu`	`bool`	`False`	Run inference on CPU even if a GPU is available.
`min_volume`	`float`	`0.0`	Minimum normalized volume for a frame to be considered.
`energy_filter_enabled`	`bool`	`True`	Apply the `min_volume` energy filter (when disabled, `min_volume` is treated as `0.0`).

Silero works with turn detection, which decides when a detected speaker has finished their turn.