Transferable speech-to-text large language model alignment module
This work addresses the challenge of efficient multimodal integration for speech-text applications, offering a simple and scalable solution that is incremental in leveraging existing models.
The paper tackles the problem of aligning speech and text modalities for tasks like spoken translation and question answering by using a one-layer module trained on a hundred-hour multitask corpus, achieving transferable alignment that works with both base and human-preference-aligned large language models.
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.