New theory explains AI hallucinations in Whisper models

A new research paper introduces the Spectral Sensitivity Theorem to explain hallucinations in large Automatic Speech Recognition (ASR) models. The theorem predicts a phase transition where models shift from signal decay to rank-1 collapse. This theory was tested on Whisper models, revealing that intermediate versions undergo structural disintegration, while larger models enter a compression-seeking attractor state that decouples them from acoustic evidence. AI

IMPACT Provides a theoretical framework to understand and potentially mitigate hallucinations in ASR models.

RANK_REASON Academic paper detailing a new theoretical framework and experimental validation. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• Kirill Borodin
• large-v3-turbo
• Self-Attention
• Spectral Sensitivity Theorem
• Tiny
• Whisper
• Whisper models

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →