Why does Whisper make up words nobody said?

Whisper makes up words because it is a neural model that generates the most likely text for the sound it hears, and it does not always stop when the sound runs out. On silence, background noise or unclear audio, it can produce fluent words that nobody spoke. Researchers call this hallucination. It is a different failure from mishearing, because the output is not a wrong version of what was said. It is text with no source in the recording.

The model can write text that has no source in the audio

Ordinary recognition errors swap or drop words that were spoken. Whisper's authors list a more serious failure among its known problems: complete hallucination, where the model outputs a transcript entirely unrelated to the actual audio. They also list repeat loops and missing words at the edges of a segment ¹. The authors report that their decoding changes help significantly, which is not the same as removing the problem.

The usual explanation is that the model writes one token after another, and a stretch with no clear speech gives it little to match, so it continues with text that looks like a transcript. That is a reasoned account, and the studies below measure the outcome rather than prove the mechanism.

How often it happens

A 2024 study, often cited as "Careless Whisper", found that roughly 1 percent of Whisper transcriptions contained entire hallucinated phrases or sentences that did not exist in the audio ². That is a small share of the total, but a single invented sentence in a medical or legal record matters.

The same study found that 38 percent of the hallucinations it analyzed included explicit harms, such as perpetuating violence, making up inaccurate associations or implying false authority ³. The study evaluated Whisper as it stood in 2023 ⁴, so newer versions and other tools may behave differently. Measure your own setup rather than rely on that figure.

Silence and background sound are known triggers

A 2025 study played non-speech sounds to Whisper and found that a set of hallucinated phrases appeared frequently ⁵. Those were artificial test inputs, so the study shows the failure mode, not how often you will meet it in daily use. It does suggest the practical rule: be most suspicious of text that appears after a long pause or over background noise.

Why Whisper can return nothing, or skip quiet speech

Whisper's own code includes a rule for silence. A segment is treated as silent when its no-speech probability is above a threshold and its average log probability is below another threshold ⁶. The defaults are 0.6 and -1.0 in the reference code. When that rule fires, the segment is dropped, so a recording that is mostly quiet can come back with little or no text.

The same rule has a cost. It cannot reliably separate very quiet speech from silence, so it can drop quiet speech too. If a dictation tool returns nothing for a passage you know you spoke, check the microphone level and the volume of your speech first.

Other tools filter silence separately. The default voice activity filter in faster-whisper removes only silences longer than 2 seconds ⁷. The same model can therefore behave differently in two apps, which is one reason an error report from one tool does not describe another.

Why the same recording can give different text

Whisper's decoder can retry a segment at higher sampling temperatures when the first attempt fails its quality checks ⁸. Sampling at a higher temperature is random by design, so a retry can produce different words from the first attempt. If you run the same recording twice and get different text, the retry path is one possible reason. Different model versions and settings can also produce different text, so note them when you compare two runs.

What to check when Whisper gives errors or fails

1. Check the audio. Confirm the microphone is the input you think it is and that the level is high enough for the speech to register. Quiet recordings are the most likely to produce silence or invented text.
2. Proofread after pauses and at the start and end. Edges and long gaps are where invented phrases and missing words most often appear.
3. Look for repeats and out-of-place text. A phrase that repeats, or a sentence you did not say, is a hallucination until proven otherwise.
4. Shorten the silence. Stop recording when you finish rather than leaving the microphone open for long stretches.
5. Test the setup with a known passage. The accuracy guide explains how to score a reference text against the output.
6. Check which version and settings you use. The guide to Whisper explains which tools use the model and how they differ.

Related reading

For how Whisper works in general, see the guide to Whisper AI. The overview of speech recognition covers the subject as a whole. The companion questions cover whether speech recognition beats humans, whether it is less accurate for some people and why it gets names and jargon wrong.