Leveraging Timestamp Information for Serialized Joint Streaming Recognition and Translation

TL;DR

This paper introduces a streaming Transformer-Transducer model that uses timestamp information for joint speech recognition and translation, enabling efficient real-time multi-language outputs with a single decoder.

Contribution

It presents a novel timestamp-based serialized output training method for joint ASR and speech translation in streaming settings, unifying multiple outputs with one model.

Findings

Effective joint recognition and translation demonstrated on multiple language pairs.

First to generate one-to-many joint outputs with a single decoder in streaming.

Improved efficiency over traditional separate systems.

Abstract

The growing need for instant spoken language transcription and translation is driven by increased global communication and cross-lingual interactions. This has made offering translations in multiple languages essential for user applications. Traditional approaches to automatic speech recognition (ASR) and speech translation (ST) have often relied on separate systems, leading to inefficiencies in computational resources, and increased synchronization complexity in real time. In this paper, we propose a streaming Transformer-Transducer (T-T) model able to jointly produce many-to-one and one-to-many transcription and translation using a single decoder. We introduce a novel method for joint token-level serialized output training based on timestamp information to effectively produce ASR and ST outputs in the streaming setting. Experiments on {it,es,de}->en prove the effectiveness of our approach, enabling the generation of one-to-many joint outputs with a single decoder for the first time.