Back to blog

Whisper Hallucinations: Why It Repeats Lines and Invents Text

Whisper hallucinations explained: why it writes “Thank you for watching” over silence or loops one line, what each fix costs, and how to find them in an SRT.

October 3, 2026
Sablate Team

Whisper hallucinations are text the model writes although nobody said it: "Thank you for watching" over silence, a subtitle credit, or one line repeated for a minute. Two settings prevent most of them — filter out non-speech audio, and stop the model conditioning on its own output — and the standard Whisper tools ship with neither on.

Below: the two kinds side by side, why each happens, what we measured when one of them hit our own app, what every fix costs, the settings for openai-whisper, faster-whisper and whisper.cpp, and how to find the survivors in a finished subtitle file.

Whisper hallucinations, side by side

Phantom phraseRepetition loopDropped speech
What you see"Thank you for watching", "Thanks for watching", a subtitle creditOne line written again and again, often with no punctuationA sentence missing from the transcript
Where it appearsSilence, music, applause, room noiseAnywhere; once it starts, it continuesQuiet openings, short interjections
What causes itNon-speech audio reaches the decoderEach window is decoded with the previous output as its promptHow a particular model handles that audio
First fixA voice-activity filter (VAD)Turn off conditioning on previous textTry a second model on the same clip
How it shows up in subtitlesAn extra cue over a pauseA cue stuck on one line, or cues with no sentence breaksA gap where a line should be

The third column is the opposite failure: speech that was skipped rather than text that was invented. It has different causes and fixes, covered in the Whisper model comparison. Everything below is about the first two.

Why does Whisper write text nobody said?

Whisper works through audio in 30-second windows, and its decoder writes text for each one. There is a no-speech check, but it is a threshold on a probability, and windows of music, applause or room tone regularly get past it and come out as words. The words it picks give the cause away. Whisper was trained on 680,000 hours of web audio paired with transcripts, and the sign-offs and subtitle credits common in that text are the usual explanation for what it writes when there is nothing to transcribe.

The numbers are not small. In a 2025 study, researchers at AGH University of Kraków ran large-v3 on non-speech audio — noise, music, environmental sound — 301,317 times. It produced text 40.3 percent of the time, and 9.1 percent of those hallucinations were loops. The most frequent outputs were "thank you", at 24.76 percent of hallucinations, and "thanks for watching", at 10.32 percent.

On real speech the rate is far lower and the damage is worse. The 2024 Careless Whisper study found that roughly 1 percent of transcriptions contained whole phrases or sentences that did not exist in the audio, that 38 percent of those included explicit harms such as perpetuating violence, inventing associations or implying false authority, and that they happened disproportionately to speakers with long pauses.

For subtitles the cost is specific. A hallucinated phrase arrives as a perfectly formatted cue over a pause, and the viewer reads it as dialogue.

Why one bad line turns into forty

The loop has a different cause. By default, Whisper decodes each window with its own previous output as a prompt, so context carries across windows: names stay spelled the same and sentences continue. The same feature lets one bad window poison the next. Once the decoder writes a line twice, the next window is primed with that line and writes it again. faster-whisper's documentation for condition_on_previous_text says it plainly: switched off, the model "becomes less prone to getting stuck in a failure loop, such as repetition looping or timestamps going out of sync".

We hit this in our own app before release, so we measured it: the same 101-second Turkish clip used in our model comparison, on an RTX 3060 Ti with float16, changing nothing but that one option.

large-v3, conditioning onlarge-v3, conditioning offmedium, conditioning onmedium, conditioning off
Sentence marks in the transcript0304730
Repeated phrases (6-grams)9000

The default wrecked large-v3 and did nothing to medium. With conditioning on, the bigger model wrote no sentence punctuation at all and produced nine repeated phrases; the smaller one, on the same audio with the same settings, was fine. In one run a single cue held this, over 0.12 seconds of audio:

[58.26 -> 58.38]  bir level oldu bir level oldu bir level oldu bir level oldu
                  bir level oldu bir level oldu bir level oldu bir level oldu

The missing punctuation matters as much as the repeats. Subtitle tools cut a transcript into cues at sentence boundaries, and a transcript without full stops gives them nothing to cut on: you get giant cues or, with a careless line splitter, one-word ones. Turning conditioning off restored the punctuation without changing the model.

One clip is not a benchmark. It does show two things: model size is no protection, and the cheapest fix is a setting.

Does turning off conditioning cost anything?

Yes. Every fix for hallucinations has a cost, and they are worth knowing before you apply them:

  1. Consistency across windows. Without the previous text as context, spelling and punctuation style can drift; the documentation warns that disabling it "may make the text inconsistent across windows". In practice: a name spelled two ways in a long file.
  2. Your initial prompt stops at the first window. Tools pass a vocabulary list — names, brands, jargon — to Whisper as its initial prompt, and with conditioning off it shapes the first window only. openai-whisper's --carry_initial_prompt True and whisper.cpp's --carry-initial-prompt re-send it for every window; in faster-whisper the hotwords option does that job.
  3. Repetition penalties delete real repetition. repetition_penalty and no_repeat_ngram_size would suppress a loop, but they cannot tell a loop from a speaker who repeats themselves. Our test clip has exactly that: one sentence said twice in a row, on purpose. A penalty that removes it is deleting dialogue.
  4. A voice-activity filter can clip quiet speech. VAD decides what counts as speech before Whisper hears it, and soft openings and band-limited audio such as phone calls are where it misses words. AutoSubs added a switch in September 2026 to turn its silence skipping off for exactly that case.
  5. hallucination_silence_threshold needs word timestamps, and it is not a cure. It skips long silences around a suspected hallucination. On our clip it made no measurable difference: the line we suspected of being invented turned out to be real speech.

None of these comes close to the cost of a loop. A misspelled name is one find-and-replace; a repetition loop is a passage you transcribe again.

How to stop Whisper hallucinating, tool by tool

openai-whisperfaster-whisperwhisper.cppSablate
Voice-activity filterNone built invad_filter=True, on by default only in the batched pipeline--vad with a Silero modelAlways on
Stop conditioning on previous text--condition_on_previous_text Falsecondition_on_previous_text=False-mc 0Always off
Skip silence after a suspected hallucination--hallucination_silence_thresholdhallucination_silence_thresholdNo equivalentNo
Re-send the prompt in every window--carry_initial_prompt Truehotwords--carry-initial-promptNo
You can change these settingsYesYesYesNo

The last three rows are Sablate's honest weak spot: both fixes are always on, and none of these settings is exposed. If you want to tune the filter or carry a prompt across windows, the open tools let you. The same two settings apply wherever Whisper feeds another program, for example when you make subtitles for DaVinci Resolve with the command line or a plugin.

For openai-whisper on an interview file:

whisper talk.mp4 --model turbo --output_format srt --condition_on_previous_text False

openai-whisper has no voice-activity filter of its own, so this command fixes the loop, not the phantom phrases; --hallucination_silence_threshold with --word_timestamps True is its partial answer to those. whisper.cpp has both fixes, and its command-line tool expects 16-bit WAV:

ffmpeg -i talk.mp4 -ar 16000 -ac 1 -c:a pcm_s16le talk.wav
./build/bin/whisper-cli -m models/ggml-large-v3-turbo.bin -f talk.wav -osrt -mc 0 --vad -vm models/ggml-silero-v6.2.0.bin

-mc 0 keeps no text context between windows, and the Silero model comes from the repository's download-vad-model.sh script. In Python, faster-whisper takes both as arguments:

from faster_whisper import WhisperModel

model = WhisperModel("large-v3-turbo")
segments, info = model.transcribe("talk.mp4", vad_filter=True, condition_on_previous_text=False)

How do you find hallucinations in a finished SRT?

No setting removes every hallucination, so check the file. Three commands catch most of the survivors, on macOS and Linux or in Git Bash and WSL on Windows:

grep -n -i -E "thank(s| you) for watching|amara\.org|subtitles by" talk.srt
awk '{ sub(/\r$/, "") } NF && !/-->/ && !/^[0-9]+$/ { if ($0 == prev) print NR": "$0; prev = $0 }' talk.srt
perl -ne 'print "$.: $_" if /(\S+(?: \S+){1,4}) \1 \1/' talk.srt

The first finds stock sign-offs and subtitle credits. The second prints any subtitle line identical to the one before it, which is how a loop usually looks once it has been cut into cues. The third catches a phrase of two to five words repeated three times inside one line.

Treat every hit as a question, not a verdict. People do say "thank you", and they do repeat themselves. Play the cue against the audio, then keep it or delete it. Phantom lines over long pauses and music match no pattern; the only check for those is to read the cues that sit over sections with no speech.

When to use which fix

"I'm subtitling talks, interviews or lectures." Turn the voice-activity filter on and conditioning off. That pair is the sensible default for any speech you plan to subtitle.

"The recording has long silences or a music bed." The filter matters most here. Read the cues over the music anyway.

"It's a phone call or a quiet, band-limited recording." Run it once with the filter and once without, and compare what is missing. A filter that drops quiet words is worse than a stray "thank you".

"I rely on an initial prompt for names." Turn conditioning off and re-send the prompt for every window with the flags above, or fix the names afterwards with find and replace.

"Lines still repeat after both fixes." Try another model on the same file before reaching for a repetition penalty; our clip looped on large-v3 and not on medium. If you still need no_repeat_ngram_size, listen through for lines it may have removed.

How Sablate handles it

Both fixes are always on in Sablate: a voice-activity filter in front of the model, and no conditioning on previous text. Sablate runs Whisper on your own machine, and we switched conditioning off after the measurement above. We also fixed the line splitter that turned a punctuation-free transcript into one-word cues.

Two consequences are worth knowing. The custom vocabulary under Defaults → Custom vocabulary reaches Whisper as its initial prompt, so with conditioning off it shapes only the first window, roughly the first 30 seconds of speech; names later in a long recording are a job for Find & replace in the editor, which works on one language at a time. And because the settings are fixed, there is nothing to tune. If a recording needs the filter switched off, use one of the open tools above.

In the editor you can step through the cues with the video playing, which is the quickest way to judge the hits from the commands above, and the export writes the corrected subtitles as .srt, .vtt or a burned-in video. Fast and Balanced run on the free plan; Accurate and Best are part of Pro. Download Sablate for Windows or Mac and try it on your own footage, or see how it compares with the other local tools in the offline subtitle generator comparison.

The short version

Whisper hallucinates in two ways: it writes stock phrases over audio with no speech, and it loops when one bad window feeds the next. A voice-activity filter fixes the first, turning off condition_on_previous_text fixes the second, and the standard tools ship with neither on. Both fixes cost little next to a loop. A bigger model is not a safer one. And whatever you change, read the finished subtitle file before anyone else does — the guide to making SRT files has the checks for the format itself.

Frequently asked questions

What are Whisper hallucinations?

Text that Whisper writes although nobody said it. The two common forms are stock phrases over silence or music, such as “Thank you for watching” or a subtitle credit, and repetition loops, where one line is written again and again. Both look like valid output, so they survive into subtitles unless someone checks the file.

Why does Whisper write “Thank you for watching”?

Because audio without speech still tends to come out as text. Whisper decodes 30-second windows, and when a window holds only music or room noise it often falls back on phrases common in its training transcripts. A 2025 study found large-v3 produced text for 40.3 percent of non-speech clips, with “thank you” and “thanks for watching” the most frequent.

Why does Whisper repeat the same line over and over?

Because by default each window is decoded with the model's own previous output as context. Once one window goes wrong, the next is primed with the mistake and repeats it. Turning off condition_on_previous_text breaks the chain. On our 101-second test clip, that took large-v3 from nine repeated phrases and no sentence punctuation to no repeats and thirty sentence marks.

How do I stop Whisper from hallucinating?

Filter out the non-speech audio and stop the model feeding on its own output. In faster-whisper that means vad_filter=True and condition_on_previous_text=False; in whisper.cpp, --vad with a Silero model and -mc 0; in the openai-whisper command line, --condition_on_previous_text False. Then check the finished file, because no setting removes every hallucination.

Does turning off condition_on_previous_text make transcripts worse?

Slightly, in specific ways. Without the previous text as context, spelling can drift between windows, and an initial prompt with names or jargon only shapes the first window unless the tool re-sends it every time. That costs a few find-and-replace edits. A repetition loop costs the whole passage, which is why turning it off is the better trade.

Is a bigger Whisper model less likely to hallucinate?

No. In our test on a 101-second Turkish clip, large-v3 fell into a repetition loop under the default settings while medium, on the same audio with the same settings, did not. Bigger models are usually more accurate on clear speech, but size gives no protection against loops or phantom phrases, so apply the same fixes to every model.