SpecAugmentTransform<T>
Models & Types · AiDotNet.Data.Transforms
Applies SpecAugment data augmentation (Park et al., 2019) to spectrogram tensors. Performs time masking and frequency masking to improve model robustness.
Implements: ITransform<Tensor<T>, Tensor<T>>
For Beginners
SpecAugment is a simple but effective augmentation for audio models. It randomly “erases” parts of the spectrogram during training to prevent overfitting.
How It Works
SpecAugment applies two types of masking to spectrograms:
- Frequency masking: Masks consecutive frequency bins (vertical stripes).
- Time masking: Masks consecutive time frames (horizontal stripes).
Reference: Park, D.S. et al. “SpecAugment: A Simple Data Augmentation Method for ASR.” Interspeech 2019.
Constructors
| Constructor | Summary |
|---|---|
SpecAugmentTransform(int freqMaskParam, int timeMaskParam, int numFreqMasks, int numTimeMasks, int? seed) |
Creates a new SpecAugment transform. |
Methods
| Method | Summary |
|---|---|
Tensor<T> Apply(Tensor<T> input) |
Applies SpecAugment masking to a spectrogram tensor. |