English

Detecting Hallucinations in SpeechLLMs at Inference Time Using Attention Maps

Computation and Language 2026-04-22 v1 Artificial Intelligence Machine Learning

Abstract

Hallucinations in Speech Large Language Models (SpeechLLMs) pose significant risks, yet existing detection methods typically rely on gold-standard outputs that are costly or impractical to obtain. Moreover, hallucination detection methods developed for text-based LLMs do not directly capture audio-specific signals. We investigate four attention-derived metrics: AUDIORATIO, AUDIOCONSISTENCY, AUDIOENTROPY, and TEXTENTROPY, designed to capture pathological attention patterns associated with hallucination, and train lightweight logistic regression classifiers on these features for efficient inference-time detection. Across automatic speech recognition and speech-to-text translation tasks, evaluations on Qwen-2-Audio and Voxtral-3B show that our approach outperforms uncertainty-based and prior attention-based baselines on in-domain data, achieving improvements of up to +0.23 PR-AUC, and generalises to out-of-domain ASR settings. We further find that strong performance can be achieved with approximately 100 attention heads, improving out-of-domain generalisation compared to using all heads. While effectiveness is model-dependent and task-specific training is required, our results demonstrate that attention patterns provide a valuable tool for hallucination detection in SpeechLLMs.

Keywords

Cite

@article{arxiv.2604.19565,
  title  = {Detecting Hallucinations in SpeechLLMs at Inference Time Using Attention Maps},
  author = {Jonas Waldendorf and Bashar Awwad Shiekh Hasan and Evgenii Tsymbalov},
  journal= {arXiv preprint arXiv:2604.19565},
  year   = {2026}
}

Comments

Accepted to Findings of ACL 2026

R2 v1 2026-07-01T12:28:33.212Z