@tensamin/audio

DeepFilterNet3-based noise suppression and realtime speaking detection for LiveKit.

Features

• DeepFilterNet3 WASM noise suppression
• Realtime speaking boolean + dB level
• Automatic mute/unmute for LiveKit tracks
• Simple min/max dB speaking thresholds

Noise suppression is provided via the deepfilternet3-noise-filter package. That package is based on DeepFilterNet by Rikorose.

Installation

npm install @tensamin/audio livekit-client

Quick Start (LiveKit)

import { LocalAudioTrack } from "livekit-client";
import { attachSpeakingDetectionToTrack } from "@tensamin/audio";

const localTrack = await LocalAudioTrack.create();

const controller = await attachSpeakingDetectionToTrack(localTrack, {
  speaking: {
    minDb: -60,
    maxDb: -20,
  },
  noiseSuppression: {
    enabled: true,
  },
  muteWhenSilent: true,
});

controller.onChange((state) => {
  console.log("speaking", state.speaking);
  console.log("levelDb", state.levelDb);
});

await room.localParticipant.publishTrack(localTrack);

Configuration

All options are passed via LivekitSpeakingOptions to attachSpeakingDetectionToTrack.

Noise suppression (DeepFilterNet3)

noiseSuppression: {
  enabled?: boolean;          // default: true
  noiseReductionLevel?: number; // 0-100, default: 60
  assetConfig?: {
    cdnUrl?: string;
  };
}

Speaking detection (dB-based)

speaking: {
  minDb: number;              // e.g. -60
  maxDb: number;              // e.g. -20
  speakOnRatio?: number;      // default: 0.6
  speakOffRatio?: number;     // default: 0.3
  hangoverMs?: number;        // default: 350
  attackMs?: number;          // default: 50
  releaseMs?: number;         // default: 120
}

minDb / maxDb define the dynamic range used for level normalization. speakOnRatio and speakOffRatio (0–1) control when speech starts/stops within that range.

Output gain control

output: {
  speechGain?: number;        // default: 1.0
  silenceGain?: number;       // default: 0.0
  gainRampTime?: number;      // default: 0.015 (s)
  maxGainDb?: number;         // default: 6.0
  smoothTransitions?: boolean;// default: true
}
``+

### LiveKit mute handling

```ts
muteWhenSilent?: boolean;     // default: false

When muteWhenSilent is true, the library automatically calls track.mute() when silence is detected and track.unmute() when speech resumes (only if it muted the track itself).