FRCRN<T>

Models & Types · AiDotNet.Audio.Enhancement

FRCRN (Frequency Recurrence CRN) speech enhancement model (Zhao et al., ICASSP 2022).

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

Implements: IAudioEnhancer<T>, IGeneratedParameterRegistrar<T>

For Beginners

FRCRN processes audio frequencies in sequence (low to high), so each frequency “knows” about its neighbors. This helps it tell speech from noise because speech frequencies appear in related patterns, while noise is more random.

Usage:

How It Works

FRCRN (Alibaba DAMO Academy) uses frequency recurrence to model spectral correlations and complex spectral mapping. It won 1st place in the ICASSP 2022 DNS Challenge non-personalized track with superior noise suppression while preserving speech quality.

Constructors

Constructor Summary
FRCRN(NeuralNetworkArchitecture<T> architecture, FRCRNOptions options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer)
FRCRN(NeuralNetworkArchitecture<T> architecture, string modelPath, FRCRNOptions options)

Properties

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

Methods

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