Whisper 환각 감지 및 완화: 숨겨진 표현 조정과 희소 자동인코더
Whisper Hallucination Detection and Mitigation via Hidden Representation Steering and Sparse AutoEncoders
- 매체
- Hugging Face
- 게시
- 중요도
- 3/5
- 화제 지수
- 3
한국어 요약
Whisper ASR 모델의 환각(음성이 아닌 입력에서 생성되는 일관된 전사) 문제를 내부 표현을 통해 감지하고 완화하는 방법을 제시했다. 희소 자동인코더(SAE) 기반 조정으로 Whisper small에서 환각률을 72.63%에서 14.11%로, large-v3에서 86.88%에서 27.33%로 감소시켰으며, 음성 데이터에서의 성능 저하는 미미했다.
원문 발췌 · English
Whisper, a widely adopted ASR model, is known to suffer from hallucinations - coherent transcriptions generated for non-speech audio entirely disconnected from the input. We investigate whether hallucinations can be detected and mitigated through Whisper's internal representations. We extract audio encoder activations and evaluate two representation spaces: raw Whisper activations and Sparse AutoEncoder (SAE) latents. We show that both spaces encode linearly separable hallucination-related information, with discriminative power concentrated in a sparse feature subset and increasing toward deeper encoder layers. We propose two steering strategies: activation-space steering and SAE latent-space steering. SAE-based steering reduces hallucination rate from 72.63% to 14.11% for Whisper small and from 86.88% to 27.33% for Whisper large-v3 on the full non-speech test set, with small WER degradation on speech data, approaching the performance of fine-tuning-based methods.