Transferable speech-to-text large language model alignment module

TL;DR

This paper introduces a transferable speech-to-text alignment module leveraging LLMs and speech models, achieving effective multimodal alignment with minimal data and demonstrating potential for multimodal expansion.

Contribution

It presents a simple, effective alignment module using one layer and limited data, applicable across different models and modalities, enhancing speech-text integration.

Findings

Modal alignment achieved with one layer and 100 hours of data

Alignment capability transfers when swapping Yi-6B with Yi-6B-Chat

Linear alignment subspace is sparse, enabling multimodal extensions

Abstract

By leveraging the power of Large Language Models(LLMs) and speech foundation models, state of the art speech-text bimodal works can achieve challenging tasks like spoken translation(ST) and question answering(SQA) altogether with much simpler architectures. In this paper, we utilize the capability of Whisper encoder and pre-trained Yi-6B. Empirical results reveal that modal alignment can be achieved with one layer module and hundred hours of speech-text multitask corpus. We further swap the Yi-6B with human preferences aligned version of Yi-6B-Chat during inference, and discover that the alignment capability is applicable as well. In addition, the alignment subspace revealed by singular value decomposition(SVD) also implies linear alignment subspace is sparse, which leaves the possibility to concatenate other features like voice-print or video to expand modality.