# Hallucination Canonical page: https://flit.fyi/glossary/hallucination [Glossary](https://flit.fyi/glossary) / [How it hears you](https://flit.fyi/glossary#recognition) Updated October 4, 2026 Definition In speech recognition, a hallucination is text the model produces that nobody said: a plausible phrase invented from silence, noise, or a fragment too short to transcribe. It is the insertion error taken to its extreme. Every speech model is trained to produce text, so when given audio that contains no speech it still tries. Batch models like Whisper are the well-known case: fed a few seconds of room tone, they will sometimes emit "Thank you for watching" or a full sentence in another language, because those strings were common in training data at the ends of recordings. The output is fluent and confidently wrong. The same word is used for language models that make up facts, and the mechanism is related: a system optimized to produce likely output will produce it whether or not the input justifies it. In a dictation tool, a hallucination means text you never spoke lands in your document. ## When it happens Silence at the start or end of a recording, which is why key-up capture and voice activity detection matter. Very short utterances, a single word or a cough, where the model has too little to anchor on. Background speech from a video or a colleague, which is real speech but not yours. Long recordings, where some models lose their place and repeat a passage. Each has a known mitigation, and a tool that has not applied them will surprise you eventually. ## How good tools prevent it Run a voice activity detector first and never send non-speech to the recognizer. Trim silence at both ends. Discard results from clips below a minimum length. Watch for the model emitting text with no corresponding audio energy and drop it. On the formatting side, refuse any rewrite that adds a content word the transcript did not contain, so a formatting model cannot hallucinate either. A tool that does all of this will still occasionally mishear you; it should never invent a sentence. ## How to reduce Whisper hallucinations from the command line If you run the open-source Whisper command-line tool on recordings and it invents or repeats text, the tool has documented options aimed at exactly this. They do not remove the problem, but they address the two common forms: text invented over silence, and a passage repeated in a loop. The Whisper authors describe these checks as a workaround for noisy predictions, not a cure, so read the result against the audio after you change them. 1. Open Terminal and run the file once with default settings, for example: whisper recording.wav --model turbo. Note where the invented or repeated text appears. 2. If a passage repeats, run it again with --condition_on_previous_text False. The documentation says this can make text less consistent between windows, but the model becomes less prone to getting stuck in a failure loop. 3. If text appears over silence, add --word_timestamps True --hallucination_silence_threshold 2. With word timestamps on, this skips silent periods longer than the given number of seconds when a possible hallucination is detected. 4. Leave --no_speech_threshold at its default of 0.6 and --logprob_threshold at its default of -1.0 unless you have a reason to change them. Together they decide when a segment is treated as silent. 5. Trim long silences from the start and end of the recording in any audio editor and run it again, so the model has less non-speech to work with. 6. Compare the new transcript with the audio at the places you noted in the first step. ## Hallucination vs mishearing: how to tell them apart A mishearing is a wrong word in the place where you said a different word. It sounds like what you said, and you can see why the model chose it. A hallucination has no source in the audio at all. The Whisper paper describes the extreme case as a transcript entirely unrelated to the actual audio. The difference matters because the fixes are different. A mishearing is improved by a better microphone, clearer speech, or a dictionary entry. A hallucination is prevented by keeping silence and noise away from the model, or by rejecting output the model itself was unsure about. - Text that appears where you were silent or had stopped speaking is a hallucination. - The same phrase or sentence repeated several times in a row is a repeat loop, which the Whisper paper lists as a separate failure from hallucination. - A wrong word that sounds like the right one is a mishearing. - The first or last few words of a recording going missing is a third failure named in the Whisper paper, and it is a deletion, not a hallucination. ## How Whisper decides that its own output has failed The open-source Whisper code checks every 30-second window after decoding it. If the average log probability of the generated tokens is below -1.0, or the gzip compression ratio of the text is above 2.4, the window is treated as failed and decoded again at a higher temperature. Highly repetitive text compresses well, which is why the compression ratio catches loops. For silence, the paper notes that the probability of the no-speech token alone is not sufficient to detect a window with no speech. Combining a no-speech threshold of 0.6 with the log probability threshold of -1 makes the detection more reliable. The model card still lists hallucination and repetitive text as known limitations. In Flit Flit checks the model’s rewrite against the transcript: if a content word you said is missing from it, the rewrite is discarded and the plain transcript lands instead. Silence and non-speech never reach the recognizer in the first place. [The details are not left to a model](https://flit.fyi/#exact) - [What is decided by rule and what by model](https://flit.fyi/#exact) - [“Every word you said survives”](https://flit.fyi/features#writes) ## Questions people ask. ### Why did my dictation app write something I never said? Most likely the recognizer was given audio with no speech in it, such as silence at the end of a recording, or a very short clip, and produced a plausible phrase anyway. This is a known behavior of batch models and it is preventable with voice activity detection and silence trimming. ### Can the formatting step hallucinate too? Yes, if it is a language model rewriting your transcript freely. The safeguard is a rule that the output may remove fillers and fix punctuation but may never add a content word you did not say, and that a rewrite which breaks that rule is thrown away. ### Why does Whisper repeat the same sentence over and over? The Whisper model card says its sequence-to-sequence architecture makes it prone to generating repetitive text. In the open-source command-line tool, setting --condition_on_previous_text False makes the model less prone to getting stuck in such a loop. ### Does a bigger Whisper model stop hallucinations? Not reliably. The Whisper paper says larger models made steady progress on perception errors such as confusing similar-sounding words, but that repeat loops and hallucination are more stubborn and are a decoding problem. ### Is a hallucination the same as a transcription error? It is one kind of transcription error. Most errors are mishearings of something you did say, while a hallucination is text with no spoken source at all, and it counts as an insertion when word error rate is calculated. ## Guides that use it - [Superwhisper not working](https://flit.fyi/superwhisper-not-working) ## Related terms - [Word error rate](https://flit.fyi/glossary/word-error-rate) - [Whisper](https://flit.fyi/glossary/whisper) - [Voice activity detection](https://flit.fyi/glossary/voice-activity-detection) - [Formatting pass](https://flit.fyi/glossary/formatting-pass) ## Your voice was always faster. 7 days free, then $19 USD once. No account, no subscription, nothing uploaded. [Download for Mac](https://flit.fyi/download) [Buy now · $19 USD](https://flit.fyi/pricing#personal) macOS 14 or later · Apple silicon