End-to-end Automatic Speech Recognition and Speech Translation: Integration of Speech Foundational Models and LLMs
Nam Luu, Ondřej Bojar
TL;DR
This work proposes an end-to-end architecture that fuses pre-trained speech encoders with Large Language Models to perform ASR and ST jointly, leveraging 4-bit QLoRA fine-tuning. By compressing speech representations via a length adapter and mapping to LLM embeddings, the model generates both transcripts and translations from a single prompt. On English→German MuST-C data, the best end-to-end configuration often beats SeamlessM4T and can match cascaded Whisper+NLLB on several metrics, with up to $8\%$ improvement in $\text{COMET}^{\text{DA}}_{22}$. Nevertheless, cascaded systems remain the strongest baseline overall, highlighting the need for further data, efficiency improvements, and architectural refinements for end-to-end speech translation.
Abstract
Speech Translation (ST) is a machine translation task that involves converting speech signals from one language to the corresponding text in another language; this task has two different approaches, namely the traditional cascade and the more recent end-to-end. This paper explores a combined end-to-end architecture of pre-trained speech encoders and Large Language Models (LLMs) for performing both Automatic Speech Recognition (ASR) and ST simultaneously. Experiments with the English-to-German language pair show that our best model not only can achieve better translation results than SeamlessM4T, a large foundational end-to-end, multi-modal translation model, but can also match the performance of a cascaded system with Whisper and NLLB, with up to a score gain of 8% in $\text{COMET}^{\text{DA}}_{22}$ metric.