denoise-voice-clarity

Real-time noise suppression for the browser — a free, open-source, license-free Krisp replacement. Runs 100% client-side in an AudioWorklet + WebAssembly (DeepFilterNet 3), so no audio ever leaves the user's device and there are no per-seat fees. Works with plain WebRTC / getUserMedia and ships a thin LiveKit adapter.

▶ Try the live demo — hear it on your own mic

Toggle the filter on your microphone, record an A/B clip, and watch a live input-vs-denoised spectrum: https://rajan471.github.io/denoise-voice-clarity/ (runs entirely in your browser — no audio is uploaded). In the clip above, red is the raw mic and green is the denoised output — flip the filter on and the green collapses below the red: that gap is the noise being removed.

• Denoise — DeepFilterNet 3 neural suppression (optional build) or a passthrough reference engine.
• Clarity — high-pass → presence EQ → AGC → soft compressor, VAD-gated.
• No license server, no per-seat cost (unlike Krisp / Koala).
• Provider-agnostic — anything that gives you a MediaStreamTrack works.
• Lazy-loaded — dynamic-import keeps the multi-MB WASM out of your initial bundle.

How it compares

| | denoise-voice-clarity | Krisp | Koala (Picovoice) | RNNoise (jitsi) | |---|---|---|---|---| | License | MIT, free | Commercial, per-seat | Commercial | BSD, free | | Runs in-browser (no server) | ✅ | ✅ (LiveKit Cloud) | ✅ | ✅ | | Model | DeepFilterNet 3 | Proprietary | Proprietary | RNNoise | | Audio leaves device | No | No | No | No | | Works outside LiveKit | ✅ (any WebRTC) | Tied to vendor SDKs | ✅ | ✅ | | Voice-clarity chain (EQ/AGC) | ✅ | partial | ❌ | ❌ |

Install

npm install denoise-voice-clarity

Quick start (any WebRTC / getUserMedia)

No LiveKit required. Dynamic-import so the WASM stays out of your initial bundle:

async function enableDenoise(): Promise<MediaStream> {
  const { createDenoisedStream, isVoiceClaritySupported } =
    await import('denoise-voice-clarity');

  const mic = await navigator.mediaDevices.getUserMedia({
    audio: { noiseSuppression: false, autoGainControl: false, echoCancellation: true },
  });

  if (!isVoiceClaritySupported()) return mic; // fall back to the raw mic

  const denoised = await createDenoisedStream(mic, {
    enabled: true,
    attenuationLimitDb: 30, // higher = more removal
    presenceGainDb: 4,      // "clarity strength"
  });

  // denoised.stream / denoised.track are clean — publish them anywhere:
  //   peerConnection.addTrack(denoised.track, denoised.stream)
  //   twilioLocalAudioTrack, daily, agora, <audio>.srcObject, …
  return denoised.stream;
}

Toggle and tune at runtime, then tear down:

denoised.setEnabled(false);     // bypass (hear the raw mic)
denoised.setPresenceGainDb(6);  // more presence
await denoised.destroy();        // free the graph + AudioContext

Turn off the browser's own noiseSuppression / autoGainControl when this is active (keep echoCancellation) so you don't double-process.

LiveKit

A TrackProcessor adapter is included — attach it to a LocalAudioTrack:

import type { LocalAudioTrack } from 'livekit-client';

async function enableVoiceClarity(micTrack: LocalAudioTrack) {
  const { VoiceClarityProcessor, isVoiceClaritySupported } =
    await import('denoise-voice-clarity');
  if (!isVoiceClaritySupported()) return false; // fall back to native suppression

  const processor = new VoiceClarityProcessor({ enabled: true, presenceGainDb: 4 });
  await micTrack.setProcessor(processor);
  return true;
}

Disable native suppression in RoomOptions.audioCaptureDefaults (keep echoCancellation). Full wiring: examples/meeting-ui-integration.ts. livekit-client ^2.17 is an optional peer dependency — only needed if you use this adapter.

API

| Export | Description | |---|---| | createDenoisedStream(stream, opts?) | Denoise a MediaStream's first audio track → DenoiseHandle. | | createDenoisedTrack(track, opts?) | Denoise a single MediaStreamTrack → DenoiseHandle. | | DenoiseEngine | Low-level engine if you want to manage the graph yourself. | | VoiceClarityProcessor | LiveKit TrackProcessor adapter. | | isVoiceClaritySupported() | Capability check (AudioWorklet + streaming WASM). |

DenoiseOptions: enabled (default true), attenuationLimitDb (0–100, default 30), presenceGainDb (−12…12, default 4). A DenoiseHandle exposes track, stream, setEnabled(), setPresenceGainDb(), and destroy().

Browser support

Requires AudioWorklet and WebAssembly.compileStreaming (all modern evergreen browsers). isVoiceClaritySupported() gates it for you; fall back to the raw mic

• native suppression otherwise.

Build from source

# 1. Rust core → WASM (needs Rust + wasm-pack; see ../scripts/build-wasm.sh)
npm run build:wasm                        # small passthrough+clarity WASM
FEATURES="wasm,dfn" npm run build:wasm    # full DeepFilterNet (needs model weights)

# 2. npm package (copies WASM, bundles worklet, emits types)
npm run build                             # → dist/{index.js, voiceClarity.worklet.js, wasm/, *.d.ts}

npm publish runs prepublishOnly → build; the published package contains only dist/ + this README.

Notes / caveats

• The worklet is a self-contained classic script with the wasm-bindgen glue inlined (AudioWorklet scope can't dynamic-import). Built by scripts/build.mjs.
• Adds ~10 ms latency (one 480-sample frame at 48 kHz) — fine for calls.
• The full DeepFilterNet model WASM is large (~18 MB); always dynamic-import it.
• The TrackProcessor interface can shift across livekit-client minors — tested against ^2.17.

License: MIT.