Reducing Linguistic Hallucination in LM-Based Speech Enhancement via Noise-Invariant Acoustic-Semantic Distillation

TL;DR

This paper introduces L3-SE, a noise-invariant acoustic-semantic distillation framework that reduces linguistic hallucination in LM-based speech enhancement, especially under adverse noise conditions.

Contribution

It proposes a novel noise-invariant conditioning encoder learned via joint distillation of acoustic and semantic targets, improving linguistic consistency in speech enhancement.

Findings

Outperforms prior LM-based SE methods on linguistic metrics

Significant reduction in hallucination under low-SNR and reverberant conditions

Maintains competitive perceptual speech quality

Abstract

Language model (LM)-based speech enhancement (SE) can generate natural-sounding speech, but under severe noise it often suffers from unreliable conditioning, leading to perceptually plausible yet linguistically incorrect outputs. To address this issue, we propose L3-SE, a noise-invariant acoustic-semantic distillation framework for reducing linguistic hallucination in LM-based SE. The proposed method learns a noise-invariant conditioning encoder from noisy speech by jointly distilling two complementary clean-speech targets: an acoustic target for reconstruction fidelity and a semantic target for linguistic consistency. The resulting noise-invariant acoustic-semantic representations are used to condition a decoder-only autoregressive language model, which predicts clean acoustic tokens that are decoded into enhanced speech. To support high-quality generation, we further employ a high-fidelity codec built on learnable weighted WavLM layer representations as the discrete acoustic interface. By improving the reliability of conditioning under adverse conditions, the proposed framework substantially reduces hallucination and improves content faithfulness. Experiments show that the proposed method consistently outperforms prior LM-based speech enhancement baselines on linguistic consistency metrics, with especially clear gains under low-SNR and reverberant conditions, while maintaining competitive perceptual quality. Audio samples are available at https://max1wz.github.io/L3-SE-Demo-Page/. The complete source code will be released after the manuscript is accepted.