Get Started
Home
Topics
Search
Library
Research questionWhere in an ASR encoder does grounding fail under distribution shift, and how is that linked to hallucination?ASR may produce transcripts that are not supported by the incoming audio when environmental or speaker conditions change. The difficulty is separating the encoder stage where audio grounding degrades from the mechanisms that might turn this failure into hallucinated text.
Audio & Speech
Audio & Speech Processing
Machine Learning
Mechanistic Interpretability
Natural Language Processing
Latest papersRecent research connected to this question, newest first.The Anatomy of an ASR HallucinationEvidence comes from two independently trained Conformer-Large recognizers: one CTC and one RNN-T. Encoder-block bypass experiments identify the final encoder stage as critical across the tested shifts, but establish a mechanistic precondition for grounding failure rather than the complete origin of naturally occurring fluent hallucinations.research paper · Sep 3, 2026
Related questions
How can sampling-only diagnosis identify whether an individual LLM hallucination comes from retrieval, training, or autoregressive commitment?How can speech recognition avoid hallucinated transcripts on non-speech without degrading genuine speech?How can we diagnose audio generation and audiovisual grounding failures in text-to-audio-video systems?How can hallucination detectors verify claims and localize unsupported spans without costly separate pipelines?