MixRep: Hidden Representation Mixup for Low-Resource Speech Recognition
This work addresses the problem of improving speech recognition accuracy in low-resource settings, which is incremental as it builds on existing mixup methods.
The paper tackles low-resource automatic speech recognition by proposing MixRep, a data augmentation strategy that interpolates hidden representations in neural networks, achieving relative WER reductions of 6.5% and 6.7% on specific datasets compared to a SpecAugment baseline.
In this paper, we present MixRep, a simple and effective data augmentation strategy based on mixup for low-resource ASR. MixRep interpolates the feature dimensions of hidden representations in the neural network that can be applied to both the acoustic feature input and the output of each layer, which generalizes the previous MixSpeech method. Further, we propose to combine the mixup with a regularization along the time axis of the input, which is shown as complementary. We apply MixRep to a Conformer encoder of an E2E LAS architecture trained with a joint CTC loss. We experiment on the WSJ dataset and subsets of the SWB dataset, covering reading and telephony conversational speech. Experimental results show that MixRep consistently outperforms other regularization methods for low-resource ASR. Compared to a strong SpecAugment baseline, MixRep achieves a +6.5\% and a +6.7\% relative WER reduction on the eval92 set and the Callhome part of the eval'2000 set.