DeepFilterNet<T>

Models & Types · AiDotNet.Audio.Enhancement

DeepFilterNet - State-of-the-art deep filtering network for speech enhancement.

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

Implements: IAudioEnhancer<T>, IGeneratedParameterRegistrar<T>

For Beginners

DeepFilterNet is like having an intelligent audio engineer that can separate speech from background noise in real-time. It’s particularly effective because it processes audio the way humans perceive sound - focusing more on frequencies that matter for understanding speech.

The model works by:

  1. Converting audio to a time-frequency representation (spectrogram)
  2. Applying learned filters to suppress noise while preserving speech
  3. Reconstructing clean audio from the enhanced spectrogram

Usage:

How It Works

DeepFilterNet is a hybrid time-frequency domain model that combines:

  • ERB (Equivalent Rectangular Bandwidth) filterbank for perceptually-motivated processing
  • Deep filtering in the complex STFT domain for fine-grained enhancement
  • Efficient architecture with grouped convolutions for real-time processing

Reference: “DeepFilterNet: A Low Complexity Speech Enhancement Framework for Full-Band Audio based on Deep Filtering” by Schröter et al., ICASSP 2022

Constructors

Constructor Summary
DeepFilterNet() Creates a DeepFilterNet model with default configuration for native training mode.
DeepFilterNet(NeuralNetworkArchitecture<T> architecture, DeepFilterNetOptions options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer, ILossFunction<T> lossFunction) Creates a DeepFilterNet model for native training and inference.
DeepFilterNet(NeuralNetworkArchitecture<T> architecture, string modelPath, DeepFilterNetOptions options, OnnxModelOptions onnxOptions) Creates a DeepFilterNet model for native training and inference.

Properties

Property Summary
bool SupportsTraining { get; } Gets whether this network supports training.
double EnhancementStrength { get; set; }
int DfOrder { get; } Gets the deep filter order.
int LatencySamples { get; }
int NumChannels { get; set; }
int NumErbBands { get; } Gets the number of ERB bands used.

Methods

Method Summary
ModelMetadata<T> GetModelMetadata()
Tensor<T> Enhance(Tensor<T> audio)
Tensor<T> EnhanceWithReference(Tensor<T> audio, Tensor<T> reference)
Tensor<T> ForwardForTraining(Tensor<T> input) Training forward used by the transparent-autodiff training path (TrainWithTape): the ERB encoder → GRU → deep-filter/gain stack composed with the SAME tape-aware IEngine layer ops as inference, so gradients flow to every trainable layer…
Tensor<T> ProcessChunk(Tensor<T> audioChunk)
void EstimateNoiseProfile(Tensor<T> noiseOnlyAudio)
void ResetState()