Minimizing Modality Gap from the Input Side: Your Speech LLM Can Be a Prosody-Aware Text LLM

2026-05-07

AI Summary

This work addresses the significant modality gap between speech large language models (SLMs) and text-based large language models (TLMs), which limits SLMs’ semantic and paralinguistic understanding. To bridge this gap, the study shifts the focus of modality alignment to the input stage by introducing a unified speech encoder that jointly generates textual tokens and prosodic embeddings, thereby constructing a prosody-aware, text-like input representation that aligns SLMs more closely with TLM processing paradigms. Built upon the WhisperPro encoder and a fine-tuned LLM backbone, the proposed approach achieves highly efficient training with only approximately 1,000 hours of audio data. It establishes the current lowest modality gap across both 3B and 7B model scales and demonstrates strong performance on paralinguistic understanding tasks.

Abstract

Speech large language models (SLMs) are typically built from text large language model (TLM) checkpoints, yet they still suffer from a substantial modality gap. Prior work has mainly attempted to reduce this gap from the output side by making speech generation more text-like, but the gap remains. We argue that the key remaining bottleneck lies on the input side. We propose TextPro-SLM, an SLM that makes spoken input more closely resemble that of a prosody-aware text LLM. TextPro-SLM combines WhisperPro, a unified speech encoder that produces synchronized text tokens and prosody embeddings, with an LLM backbone trained to preserve the semantic capabilities of the original TLM while learning paralinguistic understanding. Experiments show that TextPro-SLM achieves the lowest modality gap among leading SLMs at both 3B and 7B scales, while also delivering strong overall performance on paralinguistic understanding tasks. These gains are achieved with only roughly 1,000 hours of LLM training audio, suggesting that reducing the modality gap from the input side is both effective and data-efficient.

Problem

• modality gap
• speech large language models
• input-side alignment
• prosody-aware
• text-speech integration

Innovation

• modality gap
• prosody-aware
• speech large language model
• input-side alignment
• WhisperPro