Fact · Speech recognition
The Whisper authors list complete hallucination among known failure modes: the model outputs a transcript entirely unrelated to the actual audio, along with repeat loops and missing words at segment edges.
Checked against its source on October 17, 2026
In the source's own words
“complete hallucination where the model will output ... entirely unrelated to the actual audio”
Robust Speech Recognition via Large-Scale Weak Supervision (full text), Alec Radford et al., OpenAI (arXiv), 2022, Section 6, Limitations and Future Work.
Open the original (DOI 10.48550/arXiv.2212.04356)- What it measures
- The authors say these failures come from a mix of sequence-to-sequence modelling, language models and text-to-audio alignment. They say their decoding changes help significantly but do not remove the problem.
- What it does not show
- The authors describe the failure as a known limitation, not a rate. For a measured rate on real recordings, see the 2024 study of Whisper transcripts.
Source
Cited on
- Is Whisper AI Safe? Privacy and Security ExplainedGuide
- How Accurate Is Speech Recognition? The Real NumbersGuide
- Speech to Text and LLMs: What Language Models ChangeGuide
- Human vs AI Transcription: Accuracy, Cost, SpeedGuide
- How to Transcribe an Interview: Audio to TextGuide
- History of Speech Recognition: From Audrey to WhisperGuide
- Transcribe and Translate Audio in Other LanguagesGuide
- How accurate is AI transcription?Question
- How do you use Whisper for speech to text in Home Assistant?Question
- Is Whisper a large language model?Question
- Why does Whisper make up words nobody said?Question