Preprint
Investigation of Whisper ASR Hallucinations Induced by Non-Speech Audio
Mateusz Barański et al., arXiv (accepted for IEEE ICASSP 2025), 2025 · DOI 10.48550/arXiv.2501.11378
Read the originalRead on October 17, 2026
- What it is
- Researchers played non-speech sounds to OpenAI's Whisper and recorded the text it produced. They found a set of phrases that appear frequently when no one is speaking, and showed that post-processing can remove many of them.
- How much weight it carries
- Accepted at a major speech conference and openly available. It tests Whisper on sounds it was not built for, so it shows a failure mode, not how often ordinary dictation fails. Only the abstract was read for this guide.
Facts taken from this source
- A 2025 study of Whisper found that when non-speech sounds were played to the model, a set of hallucinated phrases appeared frequently.
“we show that there exists a set of hallucinations that appear frequently” (Abstract)
Pages that rest on this source
- Transcription: How to Turn Audio and Video Into TextTopic
- How to Transcribe Audio to TextGuide
- Meeting Transcription: How It Works and What to UseGuide
- Video to Text: Transcribe Videos and YouTubeGuide
- How Accurate Is Speech Recognition? The Real NumbersGuide
- Human vs AI Transcription: Accuracy, Cost, SpeedGuide
- Transcribe Audio With Whisper, ChatGPT or GeminiGuide
- Captions and Subtitles: Formats, Rules and ToolsGuide
- Are AI summaries of voice recordings accurate?Question
- Can word error rate be greater than 100%?Question
- How do you add automatic captions to YouTube Shorts?Question
- How do you translate a video into English subtitles?Question
- How do you turn voice notes into text with AI?Question
- Why does Whisper make up words nobody said?Question