Silero VAD Provider

Despite its name, the Silero provider does not run the Silero neural model. It is an RMS-energy detector on 16-bit little-endian PCM audio whose default threshold is calibrated to approximate Silero VAD sensitivity. Loading the Silero ONNX model is not implemented: setting ModelPath (model_path) makes construction fail instead of being silently ignored. This provider requires CGO.

Choose Silero when you want detection thresholds on Silero’s 0-1 probability scale. Do not expect neural-network accuracy in noisy environments; for real Silero inference, implement voice.ActivityDetector on an ONNX Runtime binding and register it with voice.RegisterVAD. Note that it requires CGO. For a pure Go alternative with no CGO dependency, use WebRTC.

Installation

go get github.com/lookatitude/beluga-ai/v2/voice/vad/providers/silero

This package uses the cgo build tag. Ensure CGO is enabled in your build environment.

Configuration

Field	Required	Default	Description
Threshold	No	0.5	Speech probability threshold (0.0-1.0)
SampleRate	No	16000	Audio sample rate in Hz (8000 or 16000)
ModelPath	No	—	Not supported; construction returns core.ErrInvalidInput when set

Registry configuration keys:

Key	Type	Maps to
threshold	float64	Threshold
sample_rate	int	SampleRate
model_path	string	ModelPath

Basic Usage

package main

import (
    "context"
    "fmt"
    "log"

    "github.com/lookatitude/beluga-ai/v2/voice"
    _ "github.com/lookatitude/beluga-ai/v2/voice/vad/providers/silero"
)

func main() {
    vad, err := voice.NewVAD("silero", map[string]any{
        "threshold":  0.5,
        "sample_rate": 16000,
    })
    if err != nil {
        log.Fatal(err)
    }

    // audioPCM is 16-bit little-endian PCM audio data
    var audioPCM []byte // ... obtained from audio source

    result, err := vad.DetectActivity(context.Background(), audioPCM)
    if err != nil {
        log.Fatal(err)
    }

    fmt.Printf("Speech: %v, Event: %s, Confidence: %.2f\n",
        result.IsSpeech, result.EventType, result.Confidence)
}

Advanced Features

ONNX Model Loading

ONNX model loading is not implemented. Passing model_path returns an error rather than silently falling back to the energy detector:

_, err := voice.NewVAD("silero", map[string]any{
    "threshold":  0.5,
    "model_path": "/models/silero_vad.onnx",
})
// err: "silero: ModelPath is not supported: this provider is an energy-based
// approximation and cannot load the Silero ONNX model" (core.ErrInvalidInput)

Continuous Detection with State Tracking

The provider tracks speech state internally and emits transition events:

for _, chunk := range audioChunks {
    result, err := vad.DetectActivity(ctx, chunk)
    if err != nil {
        log.Fatal(err)
    }

    switch result.EventType {
    case voice.VADSpeechStart:
        fmt.Println("Speech started")
    case voice.VADSpeechEnd:
        fmt.Println("Speech ended")
    case voice.VADSilence:
        // Ongoing silence
    }
}

Threshold Tuning

Lower thresholds increase sensitivity (more false positives), while higher thresholds reduce sensitivity (more missed speech):

• 0.3 — High sensitivity, suitable for quiet environments
• 0.5 — Balanced default for general use
• 0.7 — Low sensitivity, rejects ambient noise

Direct Construction

For compile-time type safety, construct the provider directly:

import "github.com/lookatitude/beluga-ai/v2/voice/vad/providers/silero"

vad, err := silero.New(silero.Config{
    Threshold:  0.5,
    SampleRate: 16000,
})

Error Handling

result, err := vad.DetectActivity(ctx, audioPCM)
if err != nil {
    log.Printf("VAD detection failed: %v", err)
    // Handle error — audio frame may be malformed
}

Audio data must be 16-bit little-endian PCM. Frames shorter than 2 bytes return a silence result without error.