FullSubNetPlus<T>
Models & Types · AiDotNet.Audio.Enhancement
FullSubNet+ (Full-Band and Sub-Band Fusion Network Plus) for speech enhancement.
Inheritance: NeuralNetworkBase<T> → AudioNeuralNetworkBase<T> → FullSubNetPlus<T>
Implements: IAudioEnhancer<T>, IGeneratedParameterRegistrar<T>
For Beginners
FullSubNet+ is a two-part system for cleaning up noisy speech: one part looks at the big picture (all frequencies), and another focuses on fine details (small frequency groups). Together they produce cleaner audio than either alone.
Usage:
How It Works
FullSubNet+ (Chen et al., ICASSP 2022) improves upon FullSubNet by using channel-attention-based full-band models and redesigned sub-band inputs. It achieves PESQ 3.25 and STOI 0.96 on the DNS Challenge dataset at 16 kHz.
Architecture:
- Full-band model: LSTM with channel attention across all frequency bins
- Sub-band model: LSTM processing local frequency neighborhoods
- Fusion: Full-band output guides sub-band processing
- Complex mask estimation: Predicts both magnitude and phase corrections
References:
- Paper: “FullSubNet+: Channel Attention FullSubNet with Complex Spectrograms” (Chen et al., ICASSP 2022)
- Repository:
https://github.com/hit-thusz-RookieCJ/FullSubNet-plus
Constructors
| Constructor | Summary |
|---|---|
FullSubNetPlus(NeuralNetworkArchitecture<T> architecture, FullSubNetPlusOptions options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer) |
Creates a FullSubNet+ model in native training mode. |
FullSubNetPlus(NeuralNetworkArchitecture<T> architecture, string modelPath, FullSubNetPlusOptions options) |
Creates a FullSubNet+ 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) |
Enhances audio using a reference noise sample for spectral subtraction before neural enhancement. |
Tensor<T> ProcessChunk(Tensor<T> audioChunk) |
|
void EstimateNoiseProfile(Tensor<T> noiseOnlyAudio) |
|
void ResetState() |