Dynamic Alignment Mask CTC: Improved Mask-CTC with Aligned Cross Entropy
This work addresses efficiency and accuracy in speech recognition for applications requiring fast decoding, though it appears incremental as it builds on the existing Mask-CTC framework.
The paper tackled the problem of improving non-autoregressive speech recognition models by introducing Dynamic Alignment Mask CTC, which includes Aligned Cross Entropy (AXE) and Dynamic Rectification methods. The result was improved Word Error Rate (WER) performance on the WSJ dataset, as demonstrated experimentally.
Because of predicting all the target tokens in parallel, the non-autoregressive models greatly improve the decoding efficiency of speech recognition compared with traditional autoregressive models. In this work, we present dynamic alignment Mask CTC, introducing two methods: (1) Aligned Cross Entropy (AXE), finding the monotonic alignment that minimizes the cross-entropy loss through dynamic programming, (2) Dynamic Rectification, creating new training samples by replacing some masks with model predicted tokens. The AXE ignores the absolute position alignment between prediction and ground truth sentence and focuses on tokens matching in relative order. The dynamic rectification method makes the model capable of simulating the non-mask but possible wrong tokens, even if they have high confidence. Our experiments on WSJ dataset demonstrated that not only AXE loss but also the rectification method could improve the WER performance of Mask CTC.