Balancing ASR and diarization in end-to-end LLMs for multi-talker speech recognition
For researchers in multi-talker speech recognition, this work provides a practical method to balance ASR and diarization in LLMs without large-scale annotated data.
The paper proposes strategies to efficiently train an LLM-based multi-talker speech recognition system with limited real data, achieving relative improvements of 18% on AliMeeting and 24% on Aishell4 over baselines.
Multi-talker speech recognition is often addressed by combining automatic speech recognition (ASR) and speaker diarization in a pipeline system. Recently, LLM-based approaches have shown promise by jointly modeling semantic and speaker information, but they typically require large-scale multi-talker corpora that are costly to annotate. In this paper, we investigate how to efficiently train an LLM-based system with limited real-recorded data while maintaining high accuracy in speaker attribution. We propose several strategies: (1) a dual-encoder architecture to extract semantic and speaker features, (2) a feature interleaving format to merge these features as the inputs to the LLM, (3) a length-aware speaker ID loss to enhance diarization capability, and (4) an adaptive threshold strategy for ASR loss computation to mitigate hallucinations caused by speech overlaps. These strategies balance training between ASR and diarization tasks. Our system outperforms open-source baseline approaches, achieving relative improvements of 18% on the AliMeeting corpus and 24% on the Aishell4 corpus.