livekit-plugins-dtln

In-process DTLN noise suppression for Node.js — self-hosted, open-source alternative to Krisp/noisecancellation plugins. No cloud API, no per-minute billing. Works with any LiveKit deployment or standalone.

Install

npm install livekit-plugins-dtln

ONNX model weights (~4 MB) are bundled in the package — no extra downloads needed.

Quick Start

import { DTLNNoiseSuppressor } from "livekit-plugins-dtln";

const suppressor = new DTLNNoiseSuppressor({ strength: 0.7 });
await suppressor.init();

// Process 16kHz mono float32 audio
const denoised = await suppressor.processAudio(samples);

// Or raw Int16 PCM (WebRTC-style)
const denoisedPcm = await suppressor.processInt16(pcmBuffer);

// Done
await suppressor.close();

With LiveKit (@livekit/rtc-node)

Handles any sample rate (48kHz, 16kHz, etc.) and channel count automatically:

import { DTLNNoiseSuppressor } from "livekit-plugins-dtln";
import { Room, RoomEvent, AudioStream } from "@livekit/rtc-node";

const suppressor = new DTLNNoiseSuppressor({ strength: 0.7 });
await suppressor.init();

room.on(RoomEvent.TrackSubscribed, (track) => {
  const stream = new AudioStream(track, 16000, 1);

  for await (const frame of stream) {
    const denoised = await suppressor.processFrame(
      frame.data, // Int16Array — any sample rate
      frame.sampleRate, // 48000, 16000, etc.
      frame.channels, // 1 or 2
    );
    // publish or use denoised audio...
  }
});

API

new DTLNNoiseSuppressor(options?)

| Option | Type | Default | Description | | -------------- | --------- | ------- | ------------------------------------------------ | | strength | number | 0.5 | Wet/dry blend (0 = bypass, 1 = full suppression) | | model1Path | string | bundled | Path to model_1.onnx | | model2Path | string | bundled | Path to model_2.onnx | | debugLogging | boolean | false | Log spectral mask stats every ~800ms |

Methods

| Method | Input | Output | Use case | | ----------------------------- | ------------------------------- | ----------------------- | ------------------------------- | | init() | — | Promise<void> | Load ONNX models (call once) | | processAudio(samples) | Float32Array 16kHz mono | Promise<Float32Array> | Pre-resampled audio | | processInt16(pcm) | Int16Array 16kHz mono | Promise<Int16Array> | Raw PCM buffers | | processFrame(pcm, rate, ch) | Int16Array, any rate/channels | Promise<Int16Array> | LiveKit frames (auto-resamples) | | close() | — | Promise<void> | Release ONNX sessions |

Properties

• enabled — get/set boolean to bypass processing
• initialized — read-only, true after init()

How It Works

Two-stage LSTM pipeline via ONNX Runtime:

Audio → FFT → Magnitude → [LSTM Model 1: Spectral Mask] → Masked IFFT
      → [LSTM Model 2: Time-domain Refinement] → Denoised Output

• ~10 dB noise reduction on typical speech+noise
• ~3ms latency per frame on modern hardware
• Stateful LSTM — one instance per call session
• Models from DTLN (Westhausen & Meyer, Interspeech 2020)

Platform Support

| Platform | Status | | ---------------------- | ------ | | Linux x64 | ✓ | | Linux ARM64 (Graviton) | ✓ | | macOS x64/ARM | ✓ | | Windows x64 | ✓ | | Node.js 18+ | ✓ | | CommonJS (require) | ✓ | | ESM (import) | ✓ |

Performance

• CPU: ~2-4ms per 128-sample frame (8ms of audio) on x86-64
• Memory: ~50 MB (ONNX Runtime + model weights)
• Concurrency: One instance per session; safe to run many in parallel

License

MIT

Credits

• DTLN model: breizhn/DTLN (MIT)
• ONNX Runtime: microsoft/onnxruntime (MIT)
• Python version: aloware/livekit-plugins-dtln