Noisy Environment Turn Detection

Default turn detection parameters assume clean audio. In real-world deployments — contact centers with hold music, retail floors with announcements, or factory environments with machinery — background noise causes frequent false turn-end triggers that interrupt users. This guide covers configuring both heuristic and ONNX turn detection providers for noise-resistant operation, ensuring your voice agent waits for genuine speech completion before responding.

Overview

The voice/turndetection package provides heuristic and ONNX-based turn detection. In noisy settings (contact centers, retail, factory floors), default parameters often produce too many false positives. By adjusting MinSilenceDuration, Threshold, and turn-length limits, you can achieve reliable detection despite ambient noise.

Prerequisites

• Go 1.23 or later
• A voice pipeline with VAD providing silence duration
• (Optional) ONNX turn-detection model for the onnx provider

Installation

go get github.com/lookatitude/beluga-ai

Configuration

Heuristic Provider for Noisy Settings

Increase MinSilenceDuration so brief noise gaps are not treated as end-of-turn. Use WithMinTurnLength to filter out very short spurious turns:

package main

import (
    "context"
    "fmt"
    "log"
    "time"

    "github.com/lookatitude/beluga-ai/voice/turndetection"
)

func main() {
    ctx := context.Background()

    cfg := turndetection.DefaultConfig()
    detector, err := turndetection.NewProvider(ctx, "heuristic", cfg,
        turndetection.WithMinSilenceDuration(700*time.Millisecond),
        turndetection.WithMinTurnLength(20),
        turndetection.WithMaxTurnLength(8000),
        turndetection.WithSentenceEndMarkers(".!?"),
    )
    if err != nil {
        log.Fatalf("Failed to create detector: %v", err)
    }

    audio := make([]byte, 2048)
    silence := 800 * time.Millisecond
    done, err := detector.DetectTurnWithSilence(ctx, audio, silence)
    if err != nil {
        log.Fatalf("Detection failed: %v", err)
    }
    fmt.Printf("Turn detected: %v\n", done)
}

ONNX Provider with Higher Threshold

Raise Threshold to require stronger model confidence before declaring end-of-turn, reducing false positives in noise:

import "os"

detector, err := turndetection.NewProvider(ctx, "onnx", cfg,
    turndetection.WithModelPath(os.Getenv("TURN_MODEL_PATH")),
    turndetection.WithThreshold(0.6),
    turndetection.WithMinSilenceDuration(600*time.Millisecond),
    turndetection.WithMinTurnLength(15),
)
if err != nil {
    log.Fatalf("Failed to create ONNX detector: %v", err)
}

Configuration Reference

Option	Description	Default	Noisy Environment
MinSilenceDuration	Minimum silence to trigger turn end	500 ms	600-800 ms
Threshold	ONNX detection threshold (0-1)	0.5	0.55-0.65
MinTurnLength	Minimum turn length	10	15-25
MaxTurnLength	Maximum turn length	5000	8000+
SentenceEndMarkers	Heuristic sentence-end characters	.!?	Keep or extend

Troubleshooting

Too many false turn-end events in noise

Increase MinSilenceDuration (600-800 ms) and, for ONNX, Threshold (0.55-0.65). Use DetectTurnWithSilence fed by a robust VAD so silence is computed from actual speech absence rather than raw audio energy.

Missed end-of-turn in noisy segments

Ensure VAD correctly identifies silence. Avoid over-incrementing MinSilenceDuration. Consider switching to the ONNX provider if heuristic detection is insufficient. Verify audio format (sample rate, chunk size) matches provider expectations.

Provider ‘onnx’ not registered

Import the ONNX provider to trigger registration:

import _ "github.com/lookatitude/beluga-ai/voice/turndetection/providers/onnx"

Set TURN_MODEL_PATH to a valid ONNX model file path.

Advanced Topics

Production Deployment

• Use turndetection.IsRetryableError(err) and retry where appropriate
• Call turndetection.InitMetrics(meter, tracer) at startup for OpenTelemetry monitoring
• A/B test heuristic vs ONNX providers with different thresholds using metrics before rollout
• Run with real or synthetic noisy audio to validate settings against ground truth

Related Resources

• Heuristic Turn Detection Tuning
• Voice Services Overview