VAD Providers — Voice Detection

Beluga AI provides a unified voice.VAD interface for detecting speech in audio streams. VAD providers analyze audio frames and report whether speech is present, along with confidence scores and state transitions (speech start, speech end, silence).

How It Works

All VAD providers implement the same interface:

type VAD interface {
    DetectActivity(ctx context.Context, audio []byte) (ActivityResult, error)
}

The ActivityResult contains:

type ActivityResult struct {
    IsSpeech   bool         // true if speech was detected
    EventType  VADEventType // speech_start, speech_end, or silence
    Confidence float64      // detection confidence (0.0 to 1.0)
}

You can instantiate any provider two ways:

Via the registry (recommended for dynamic configuration):

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

vad, err := voice.NewVAD("silero", map[string]any{
    "threshold": 0.5,
})

Via direct construction (for compile-time type safety):

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

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

Event Types

VAD providers track state transitions between speech and silence:

Event Type	Constant	Description
Speech Start	VADSpeechStart	Transition from silence to speech
Speech End	VADSpeechEnd	Transition from speech to silence
Silence	VADSilence	No speech detected (ongoing)

Built-in Energy VAD

The voice package includes a built-in energy-threshold VAD that requires no external dependencies. It computes the RMS energy of 16-bit PCM audio and compares it against a configurable threshold:

import "github.com/lookatitude/beluga-ai/voice"

vad, err := voice.NewVAD("energy", map[string]any{
    "threshold": 1000.0,
})

Available Providers

Provider	Registry Name	Description
Energy	energy	Built-in RMS energy threshold detector
Silero	silero	ONNX model-based detection with energy fallback
WebRTC	webrtc	Energy + zero-crossing rate analysis

Provider Discovery

List all registered VAD providers at runtime:

for _, name := range voice.ListVAD() {
    fmt.Println(name)
}

Choosing a Provider

Use Case	Recommended Provider	Reason
Quick prototyping	energy	Zero dependencies, built-in
Production accuracy	silero	Neural network-based detection
Low-latency	webrtc	Lightweight dual-metric analysis
Noise filtering	webrtc	Zero-crossing rate rejects noise