DeepFilterNet3 Noise Filter for LiveKit

AI-powered noise suppression for real-time audio processing with LiveKit.

Based on the DeepFilterNet paper and implementation by Rikorose.

Installation

npm install deepfilternet3-noise-filter

Usage

Basic Audio Processing

import { DeepFilterNet3Core } from 'deepfilternet3-noise-filter';

// Create audio context
const ctx = new (window.AudioContext || window.webkitAudioContext)({ sampleRate: 48000 });

// Initialize processor
const proc = new DeepFilterNet3Core({
  sampleRate: 48000,
  noiseReductionLevel: 0
});

await proc.initialize();

// Create audio worklet node
const node = await proc.createAudioWorkletNode(ctx);

// Connect your audio stream
const src = ctx.createMediaStreamSource(stream);
const dst = ctx.createMediaStreamDestination();
src.connect(node).connect(dst);

// Adjust noise reduction level (0-100)
proc.setSuppressionLevel(50);

React Example

import React, { useRef, useEffect } from 'react';
import { DeepFilterNet3Core } from 'deepfilternet3-noise-filter';

function AudioProcessor({ stream, level = 50 }) {
  const ctxRef = useRef(null);
  const procRef = useRef(null);
  const nodeRef = useRef(null);

  useEffect(() => {
    const setupAudio = async () => {
      const ctx = new (window.AudioContext || window.webkitAudioContext)({ sampleRate: 48000 });
      ctxRef.current = ctx;

      const proc = new DeepFilterNet3Core({
        sampleRate: 48000,
        noiseReductionLevel: 0
      });

      await proc.initialize();
      procRef.current = proc;
      
      const node = await proc.createAudioWorkletNode(ctx);
      nodeRef.current = node;

      const src = ctx.createMediaStreamSource(stream);
      const dst = ctx.createMediaStreamDestination();
      src.connect(node).connect(dst);
      proc.setSuppressionLevel(level);
    };

    if (stream) {
      setupAudio();
    }

    return () => {
      if (procRef.current) {
        procRef.current.destroy();
      }
    };
  }, [stream, level]);

  return null; // This component only handles audio processing
}

No configuration needed - WebAssembly files and worker code are automatically handled!

Bundler Compatibility

This package works out-of-the-box with all modern bundlers:

• Webpack (4, 5+)
• Vite
• Rollup
• esbuild
• Parcel

Worker and worklet files are automatically inlined as blob URLs, so no webpack configuration or copy plugins are required. Just npm install and use!

LiveKit Integration

import { DeepFilterNoiseFilterProcessor } from 'deepfilternet3-noise-filter';

// Create the processor
const filter = new DeepFilterNoiseFilterProcessor({
  sampleRate: 48000,
  noiseReductionLevel: 80,
  enabled: true,
  assetConfig: {
    cdnUrl: 'https://cdn.mezon.ai/AI/models/datas/noise_suppression/deepfilternet3' // Optional: use custom CDN
  }
});

// Use with LiveKit
await audioTrack.setProcessor(filter);
await room.localParticipant.publishTrack(audioTrack);

// Control noise reduction
filter.setSuppressionLevel(60);
filter.setEnabled(false); // Disable temporarily

For a complete React example, see: DeepFilterNet3 React Example

Custom CDN Configuration

By default, the package loads WASM and model files from the bundled assets. You can optionally configure a custom CDN:

const filter = new DeepFilterNoiseFilterProcessor({
  sampleRate: 48000,
  noiseReductionLevel: 80,
  assetConfig: {
    cdnUrl: 'https://your-cdn-url.com/path/to/assets'
  }
});

Version-Specific Asset Paths

Different versions use different asset path structures:

Version <= 1.1.2:

• Assets are loaded directly from the base path
• WASM: {cdnUrl}/pkg/df_bg.wasm
• Model: {cdnUrl}/models/DeepFilterNet3_onnx.tar.gz

Version >= 1.2.0:

• Assets are loaded from a v2/ subdirectory (automatically added)
• WASM: {cdnUrl}/v2/pkg/df_bg.wasm
• Model: {cdnUrl}/v2/models/DeepFilterNet3_onnx.tar.gz
• Built with SIMD optimizations for ~20-30% better performance

Note: The v2/ prefix is added automatically by the package for version >= 1.2.0, so you don't need to include it in your cdnUrl configuration.

Build

yarn
yarn build

Outputs:

• dist/

Model source

This package is based on DeepFilterNet by Rikorose.

Original Paper:

• Schröter, H., Rosenkranz, T., Escalante-B., A.N., & Maier, A. (2022). DeepFilterNet: A Low Complexity Speech Enhancement Framework for Full-Band Audio based on Deep Filtering. ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 7407-7411.
• Paper on arXiv

The included model archive DeepFilterNet3_onnx.tar.gz is from:

• DeepFilterNet3_onnx.tar.gz

Please refer to the upstream repository for licensing and updates.

Building WASM from source (contributors)

WASM Build Configuration

The WASM module is built with optimized settings in libDF/Cargo.toml:

Rust Compiler Settings ([profile.release]):

opt-level = 3          # Maximum optimization for performance
lto = "thin"           # Thin Link-Time Optimization
codegen-units = 1      # Single codegen unit for better optimization
panic = "abort"        # Smaller binary size

WASM Optimizer Settings ([package.metadata.wasm-pack.profile.release]):

wasm-opt = [
    "-O4",                                  # Highest optimization level
    "--enable-simd",                        # Enable simd
    "--enable-bulk-memory",                 # Enable bulk memory operations
    "--enable-nontrapping-float-to-int"     # Enable non-trapping float-to-int conversions
]

Build Steps

To regenerate the WASM package and copy resources from the upstream project:

# Clone the DeepFilterNet repository
git clone https://github.com/Rikorose/DeepFilterNet/
cd DeepFilterNet/libDF

# Configure Cargo.toml with the optimization settings above
# Then build the WASM package with target features
RUSTFLAGS="-C target-feature=+simd128,+bulk-memory,+nontrapping-fptoint,+mutable-globals" \
  wasm-pack build --target web --release --features wasm

# Copy WASM files to this repo
cd ..
cp -r libDF/pkg ../mezon-noise-suppression/df3

cd ../mezon-noise-suppression
yarn build

Build Notes:

• RUSTFLAGS target features:
  • +simd128: Enable SIMD 128-bit operations for vectorized processing
  • +bulk-memory: Enable efficient bulk memory operations (memcpy, memset)
  • +nontrapping-fptoint: Enable non-trapping float-to-int conversions
  • +mutable-globals: Enable mutable global variables
• Version <= 1.1.2: Built without advanced WebAssembly features
• Version >= 1.2.0: Built with full WebAssembly feature set for optimal performance
• Ensure wasm-pack is installed: cargo install wasm-pack
• Browser requirements: Chrome 91+, Firefox 89+, Safari 16.4+ (for SIMD support)

License

This project is dual-licensed under either:

• Apache License, Version 2.0 (LICENSE-APACHE or http://www.apache.org/licenses/LICENSE-2.0)
• MIT License (LICENSE-MIT or http://opensource.org/licenses/MIT)

at your option.

This licensing follows the upstream DeepFilterNet project by Hendrik Schröter, which uses the same dual-license approach. The DeepFilterNet WASM binaries and models included in this package are governed by the same licenses.

Contribution

Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in the work by you, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.