Align-Refine: Non-Autoregressive Speech Recognition via Iterative Realignment

TL;DR

Align-Refine introduces an iterative realignment approach for non-autoregressive speech recognition, significantly improving accuracy and speed by refining latent alignments rather than output sequences.

Contribution

The paper presents a novel iterative realignment method that refines latent alignments in non-autoregressive models, enabling length-changing edits and outperforming previous models.

Findings

Outperforms Imputer and Mask-CTC on WSJ and LibriSpeech.

Matches autoregressive baseline at 1/14th real-time factor.

Achieves 9.0% WER on LibriSpeech test-other without an LM.

Abstract

Non-autoregressive models greatly improve decoding speed over typical sequence-to-sequence models, but suffer from degraded performance. Infilling and iterative refinement models make up some of this gap by editing the outputs of a non-autoregressive model, but are constrained in the edits that they can make. We propose iterative realignment, where refinements occur over latent alignments rather than output sequence space. We demonstrate this in speech recognition with Align-Refine, an end-to-end Transformer-based model which refines connectionist temporal classification (CTC) alignments to allow length-changing insertions and deletions. Align-Refine outperforms Imputer and Mask-CTC, matching an autoregressive baseline on WSJ at 1/14th the real-time factor and attaining a LibriSpeech test-other WER of 9.0% without an LM. Our model is strong even in one iteration with a shallower decoder.