CTC-based Non-autoregressive Textless Speech-to-Speech Translation

TL;DR

This paper explores CTC-based non-autoregressive models for speech-to-speech translation, achieving comparable quality to autoregressive models while significantly increasing decoding speed through advanced training techniques.

Contribution

It introduces a CTC-based NAR approach for S2ST, combining pretraining, knowledge distillation, and novel training methods to match AR model quality with much faster decoding.

Findings

Achieves up to 26.81× decoding speedup.

Attains translation quality comparable to AR models.

Demonstrates effectiveness of advanced NAR training techniques.

Abstract

Direct speech-to-speech translation (S2ST) has achieved impressive translation quality, but it often faces the challenge of slow decoding due to the considerable length of speech sequences. Recently, some research has turned to non-autoregressive (NAR) models to expedite decoding, yet the translation quality typically lags behind autoregressive (AR) models significantly. In this paper, we investigate the performance of CTC-based NAR models in S2ST, as these models have shown impressive results in machine translation. Experimental results demonstrate that by combining pretraining, knowledge distillation, and advanced NAR training techniques such as glancing training and non-monotonic latent alignments, CTC-based NAR models achieve translation quality comparable to the AR model, while preserving up to 26.81$\times$ decoding speedup.