NAIST Simultaneous Speech Translation System for IWSLT 2024

TL;DR

This paper presents NAIST's multilingual end-to-end speech translation system for IWSLT 2024, combining pre-trained models and innovative policies to enhance speech-to-text and speech-to-speech translation quality.

Contribution

The paper introduces a novel multilingual speech translation system integrating HuBERT and mBART with new decoding policies and an improved incremental TTS module using Transformer architecture.

Findings

LA decoding policy outperforms AlignAtt in previous models

Enhanced TTS module with Transformer architecture improves system performance

Cascade of speech-to-text and incremental TTS achieves effective speech translation

Abstract

This paper describes NAIST's submission to the simultaneous track of the IWSLT 2024 Evaluation Campaign: English-to-{German, Japanese, Chinese} speech-to-text translation and English-to-Japanese speech-to-speech translation. We develop a multilingual end-to-end speech-to-text translation model combining two pre-trained language models, HuBERT and mBART. We trained this model with two decoding policies, Local Agreement (LA) and AlignAtt. The submitted models employ the LA policy because it outperformed the AlignAtt policy in previous models. Our speech-to-speech translation method is a cascade of the above speech-to-text model and an incremental text-to-speech (TTS) module that incorporates a phoneme estimation model, a parallel acoustic model, and a parallel WaveGAN vocoder. We improved our incremental TTS by applying the Transformer architecture with the AlignAtt policy for the estimation model. The results show that our upgraded TTS module contributed to improving the system performance.