SpecAugment: A Simple Data Augmentation Method for Automatic Speech Recognition

SpecAugment is a data augmentation method for automatic speech recognition that operates directly on the filter bank feature inputs of a neural network rather than on raw audio. Its augmentation policy combines three operations: warping the features along the time axis, masking contiguous blocks of frequency channels, and masking contiguous blocks of time steps. The authors apply it to Listen, Attend and Spell (LAS) networks trained for end-to-end speech recognition.

On the LibriSpeech 960h and Switchboard 300h benchmarks, SpecAugment set new state-of-the-art results, outperforming all prior work including sophisticated hybrid systems. On LibriSpeech test-other it reached 6.8% word error rate without any language model and 5.8% with shallow fusion, versus the previous best hybrid system's 7.5%; on the Switchboard/CallHome test set it achieved 7.2%/14.6% (6.8%/14.1% with fusion) against a prior 8.3%/17.3%. These gains came from a low-cost augmentation applied only to input features, showing a simple method can surpass more complex prior systems.