How do I check for hallucinations quickly?

Play the silent and musical passages rather than reading them. Any text over silence is suspect by definition, and those spans are quick to find in a timed transcript.

Is a hallucination the same as a high word error rate?

No. WER counts insertions, deletions and substitutions equally. A hallucination is a block of insertions that can leave overall WER looking acceptable while the transcript contains invented statements.

Can hallucinations be prevented?

Reduced, not eliminated. Trimming leading and trailing silence, and supplying clean audio, both cut the frequency considerably.

Does trimming silence reduce hallucinations?

It helps noticeably. Leading and trailing silence is where invented text most often appears, so removing it before processing cuts the most common case.

Definition

What Is an ASR Hallucination?

VideoText editorial team. Updated September 13, 2026.

Direct definition

In speech recognition, a hallucination is invented text that does not correspond to the audio — for example, repeated phrases, leftover prompt language, or fluent sentences during silence or music.

The sections below explain how hallucination is used in transcription and subtitle work, what people often mix up, and which VideoText workflow applies when the next step is a file or review pass.

Key takeaways

Hallucinations are insertions, often fluent and therefore dangerous.
They are not the same as a misheard homophone.
Always spot-check silence, music beds, and the start of files.

How to treat them in QA

Delete invented spans. Do not “lightly edit” a hallucination into something that sounds plausible. If the audio is inaudible, use the style guide’s inaudible mark.

Where hallucinations cluster

They appear most often where there is little or no speech for the model to work from: silence at the head or tail of a file, music beds, applause, sustained background noise, and long pauses between speakers. With nothing to transcribe, some models emit fluent text anyway.

A recognisable pattern is repetition — the same phrase looping several times — and another is text that belongs to the training data rather than the recording, such as a stray subtitle credit or a thank-you line.

Why they are more dangerous than ordinary errors

A misheard word usually looks wrong in context, so a reviewer catches it. A hallucination is grammatical, plausible and confident, which means it survives a skim read and can be quoted downstream as though it were said.

For legal, medical and journalistic work this is the failure mode that matters most, because the resulting transcript is not merely inaccurate but asserts something that never happened.

Frequently asked questions

How do I check for hallucinations quickly?: Play the silent and musical passages rather than reading them. Any text over silence is suspect by definition, and those spans are quick to find in a timed transcript.
Is a hallucination the same as a high word error rate?: No. WER counts insertions, deletions and substitutions equally. A hallucination is a block of insertions that can leave overall WER looking acceptable while the transcript contains invented statements.
Can hallucinations be prevented?: Reduced, not eliminated. Trimming leading and trailing silence, and supplying clean audio, both cut the frequency considerably.
Does trimming silence reduce hallucinations?: It helps noticeably. Leading and trailing silence is where invented text most often appears, so removing it before processing cuts the most common case.

Related terms

Sources

OpenAI: Whisper paper (robustness and error analysis context)