From TOWER to SPIRE: Adding the Speech Modality to a Text-Only LLM

Abstract

Large language models (LLMs) have shown remarkable performance and generalization capabilities across multiple languages and tasks, making them very attractive targets for multi-modality integration (e.g., images or speech). In this work, we extend an existing LLM to the speech modality via speech discretization and continued pre-training. In particular, we are interested in multilingual LLMs, such as TOWER, as their pre-training setting allows us to treat discretized speech input as an additional translation language. The resulting open-source model, SPIRE, is able to transcribe and translate English speech input while maintaining TOWER's original performance on translation-related tasks, showcasing that discretized speech input integration as an additional language is feasible during LLM adaptation. We make our code and models available to the community.