NeuralNoiseReducer<T>

Models & Types · AiDotNet.Audio.Enhancement

Neural network-based noise reducer for high-quality audio enhancement.

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

Implements: IAudioEnhancer<T>, IGeneratedParameterRegistrar<T>

For Beginners

This is like a “magic eraser” for audio noise!

How it works:

  1. Converts audio to a spectrogram (picture of sound)
  2. Neural network learns to identify and remove noise patterns
  3. Converts cleaned spectrogram back to audio

Key features:

  • Works on any type of noise (AC hum, fan noise, traffic, etc.)
  • Preserves speech/music quality while removing noise
  • Can be trained on your specific noise conditions
  • Supports real-time streaming for live applications

Use cases:

  • Podcast/video production (remove background noise)
  • Voice calls (improve speech clarity)
  • Music restoration (remove hiss/crackle from old recordings)
  • Hearing aids (enhance speech in noisy environments)

Two modes of operation:

  1. ONNX Mode: Load a pre-trained model for instant use
  2. Native Mode: Train your own model on custom data

How It Works

This model uses an encoder-bottleneck-decoder architecture inspired by U-Net to learn the mapping from noisy audio to clean audio. It operates in the time-frequency domain using STFT for analysis and synthesis. Note: The current implementation uses a simplified dense decoder; a full U-Net with transposed convolutions and skip connections is planned for future versions.

Example

var noisyAudioTensor = Tensor<float>.CreateRandom(1, 16000);
// Create a neural noise reducer with default settings
var reducer = new NeuralNoiseReducer<float>();
            
// Or with custom architecture
var architecture = new NeuralNetworkArchitecture<float>(
    inputType: InputType.OneDimensional,
    taskType: NeuralNetworkTaskType.Regression,
    inputSize: 257,
    outputSize: 257);
var customReducer = new NeuralNoiseReducer<float>(architecture);
            
// Enhance noisy audio
Tensor<float> cleanAudio = reducer.Enhance(noisyAudioTensor);

Constructors

Constructor Summary
NeuralNoiseReducer() Creates a NeuralNoiseReducer with default configuration for native training mode.
NeuralNoiseReducer(NeuralNetworkArchitecture<T> architecture, NeuralNoiseReducerOptions options, ILossFunction<T> lossFunction) Creates a NeuralNoiseReducer in native training mode.
NeuralNoiseReducer(NeuralNetworkArchitecture<T> architecture, string modelPath, NeuralNoiseReducerOptions options) Creates a NeuralNoiseReducer in native training mode.

Properties

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

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)
void ResetEnhancerState() Resets the enhancer streaming state.