DemucsNoise<T>

Models & Types · AiDotNet.Audio.Effects

Demucs for Noise - real-time noise suppression using the Demucs architecture (Defossez et al., 2020, Meta).

Inheritance: NeuralNetworkBase<T> → AudioNeuralNetworkBase<T> → DemucsNoise<T>

Implements: IAudioEnhancer<T>

For Beginners

Demucs for Noise works like a music separator, but instead of separating instruments, it separates clean speech from background noise. Feed it a noisy recording and it outputs just the clean speech.

How it works:

  1. Encoder: Progressively compresses audio into a compact representation
  2. LSTM bottleneck: Captures temporal patterns in the compressed audio
  3. Decoder: Reconstructs clean audio with skip connections from the encoder

Usage:

How It Works

Demucs for Noise adapts the Demucs U-Net encoder-decoder architecture for real-time speech denoising. Operating in the time domain with skip connections and LSTM bottleneck, it achieves high-quality noise removal at 40ms latency. It was developed at Meta for real-time communication applications.

Constructors

Constructor Summary
DemucsNoise(NeuralNetworkArchitecture<T> architecture, DemucsNoiseOptions options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer) Creates a Demucs Noise model in native training mode.
DemucsNoise(NeuralNetworkArchitecture<T> architecture, string modelPath, DemucsNoiseOptions options) Creates a Demucs Noise model in native training mode.

Properties

Property Summary
double EnhancementStrength { get; set; }
int LatencySamples { get; }
int NumChannels { get; }

Methods

Method Summary
ModelMetadata<T> GetModelMetadata()
Tensor<T> Enhance(Tensor<T> audio)
Tensor<T> EnhanceWithReference(Tensor<T> audio, Tensor<T> reference)
Tensor<T> ProcessChunk(Tensor<T> audioChunk)
void EstimateNoiseProfile(Tensor<T> noiseOnlyAudio)