Soundwave: Less is More for Speech-Text Alignment in LLMs

TL;DR

Soundwave introduces a novel architecture and training strategy for speech-text alignment in LLMs, achieving high performance with significantly less training data by addressing representation and sequence length issues.

Contribution

The paper presents a new model and training approach that effectively reduces data requirements for speech-text alignment in LLMs, outperforming existing models.

Findings

Outperforms Qwen2-Audio in speech translation and AIR-Bench tasks

Uses only one-fiftieth of the training data compared to previous models

Retains conversational intelligence despite reduced training data

Abstract

Existing end-to-end speech large language models (LLMs) usually rely on large-scale annotated data for training, while data-efficient training has not been discussed in depth. We focus on two fundamental problems between speech and text: the representation space gap and sequence length inconsistency. We propose Soundwave, which utilizes an efficient training strategy and a novel architecture to address these issues. Results show that Soundwave outperforms the advanced Qwen2-Audio in speech translation and AIR-Bench speech tasks, using only one-fiftieth of the training data. Further analysis shows that Soundwave still retains its intelligence during conversation. The project is available at https://github.com/FreedomIntelligence/Soundwave.