Soundwave: Less is More for Speech-Text Alignment in LLMs
This work addresses the need for more efficient training methods in speech-text alignment for LLMs, offering a domain-specific solution that reduces data requirements while maintaining performance.
The paper tackles the problem of data-efficient training for speech-text alignment in LLMs by addressing representation space gaps and sequence length inconsistencies, achieving superior performance on speech translation and AIR-Bench tasks with only 1/50th of the training data compared to Qwen2-Audio.
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.