Two-Stage Acoustic Adaptation with Gated Cross-Attention Adapters for LLM-Based Multi-Talker Speech Recognition

TL;DR

This paper enhances multi-talker speech recognition with large language models by explicitly injecting talker-aware acoustic evidence through a two-stage, parameter-efficient adaptation framework, leading to improved performance especially in three-talker scenarios.

Contribution

It introduces a novel two-stage acoustic adaptation method using gated cross-attention adapters and LoRA updates for robust multi-talker ASR with LLMs.

Findings

Significant performance gains on Libri2Mix/Libri3Mix datasets.

Large improvements in three-talker mixture recognition.

Effective robustness under noisy conditions.

Abstract

Large Language Models (LLMs) are strong decoders for Serialized Output Training (SOT) in two-talker Automatic Speech Recognition (ASR), yet their performance degrades substantially in challenging conditions such as three-talker mixtures. A key limitation is that current systems inject acoustic evidence only through a projected prefix, which can be lossy and imperfectly aligned with the LLM input space, providing insufficient fine-grained grounding during decoding. Addressing this limitation is crucial for robust multi-talker ASR, especially in three-talker mixtures. This paper improves LLM-based multi-talker ASR by explicitly injecting talker-aware acoustic evidence into the decoder. We first revisit Connectionist Temporal Classification (CTC)-derived prefix prompting and compare three variants with increasing acoustic content. The CTC information is obtained using the serialized CTC proposed in our previous works. While acoustic-enriched prompts outperform the SOT-only baseline, prefix-only conditioning remains inadequate for three-talker mixtures. We therefore propose a lightweight gated residual cross-attention adapter and design a two-stage acoustic adaptation framework based on low-rank updates (LoRA). In Stage 1, we insert gated cross-attention adapters after the self-attention sub-layer to stably inject acoustic embeddings as external memory. In Stage 2, we refine both the cross-attention adapters and the pretrained LLM's self-attention projections using parameter-efficient LoRA, improving robustness for large backbones under limited data; the learned updates are merged into the base weights for inference. Experiments on Libri2Mix/Libri3Mix under clean and noisy conditions show consistent gains, with particularly large improvements in three-talker settings.