SpecAugment
A speech data augmentation method proposed by Google in 2019 that applies random time masking and frequency masking to mel spectrograms, forcing the model to make predictions under incomplete information and thereby improving ASR model robustness to noise and missing information.
Timeline (last 90 days)
The two categories of data augmentation methods compared in the study are methods that modify existing audio (speed perturbation and SpecAugment) and methods that modify text transcriptions and synthesize audio accordingly.
SpecAugment yielded the most stable and reliable improvements on aggregate metrics, but increased tonal and palatalization errors.
All Facts (2)
The two categories of data augmentation methods compared in the study are methods that modify existing audio (speed perturbation and SpecAugment) and methods that modify text transcriptions and synthesize audio accordingly.
50%UnverifiedSpecAugment yielded the most stable and reliable improvements on aggregate metrics, but increased tonal and palatalization errors.
50%