On this page
A speech model has no way to say nothing. It takes a stretch of sound and returns the most likely sequence of words for it, so when the sound holds no speech, it still returns words. That is transcription hallucination: words nobody said, arriving as fluent, correctly punctuated text. It takes four shapes, and knowing them is most of the cure: a polite phrase dropped over silence, one sentence repeated a dozen times, invented lines over music, and a plausible wrong word where the speech was buried. The fix is not a smarter model. It is refusing to let the model see audio with no speech in it, which is what a voice activity detector does, and on FastScribe it runs on every file before a word is decoded.
What transcription hallucination is, and why the invented text reads so well
The model is not lying, it is completing. Speech recognition scores whole sequences rather than matching sounds to words one at a time, weighing what the audio sounds like against what a plausible sentence looks like. That second half is why a transcript comes back with grammar and punctuation instead of phonetic soup, and it is why the failures read so well: when the audio carries no usable evidence, the acoustic side contributes almost nothing and the language side fills the gap alone.
That is the whole problem. A hallucinated line has the same fluency, punctuation and confidence as a correct one, and nothing marks it. You do not find it by reading carefully. You find it by knowing where it collects. Training data leaves fingerprints too, which is why models trained on internet video put "Thanks for watching" at the end of a research interview.
The four shapes it takes, and how to find each one
Three of the four collect in predictable places, which is what makes them cheap to catch: a targeted check of the right thirty seconds beats a full proofread of the whole transcript.
A polite phrase over silence. "Thank you." "Thanks for watching." It lands wherever the recording has a gap: before anyone starts, after everyone stops, and in long pauses. Check the first and last thirty seconds of every transcript. Commonest form, cheapest to catch.
A repeated line. The model sticks over sustained noise and emits the same sentence eight or twelve times. Skim the shape of the page rather than its content, looking for a block where every line is identical. In an SRT or VTT export it is unmissable: one cue repeating while the timestamps advance.
Invented speech over music. Sung vocals are speech-like enough to be decoded and unlike enough that what comes back is wrong. If your file has a music bed under it, expect invented lines rather than lyrics. Wrong job for the tool, not a setting to change.
A confident wrong word in buried audio. The hard one, and not invention from nothing: invention from a phrase the model could not resolve. A name it has never met, a number under crosstalk, a "not" swallowed by a cough. It reads perfectly and its meaning has changed. No gate removes this shape, which is why a passage you intend to quote gets checked against the audio by a person. How to edit and clean up a transcript covers that pass.
Want to try it now? Upload a file to FastScribe. One file a week is free, no signup.
Why silence is the worst input you can give a speech model
Silence is the worst input because it carries no acoustic evidence at all, which leaves the language side of the model to decide the output entirely on its own, with nothing to check it against.
True digital silence measures around -91 dB of mean volume. A room with a recorder running and nobody talking sits far above that: air conditioning, mains hum, somebody's sleeve on the table. It is that near-silence, not an empty file, that produces most invented text: enough sound to look like something, not enough to be anything.
Leave a recorder running for ten minutes before an interview starts and you have manufactured ten minutes of the hardest possible input, at the front of the file where you are least likely to reread it. Trim dead air off both ends before you upload. It removes a whole class of error, and on any plan with a length cap it buys back minutes. How to prepare audio so it transcribes accurately covers the rest of the recording side.
How FastScribe handles it
A voice activity detector runs before transcription and hands the model only the stretches that hold speech. Non-speech audio never reaches the decoder, so there is nothing there to complete. That is what keeps "Thank you." out of a transcript of a mostly quiet room. The detector is a separate speech-trained model, not a setting on the first one, and the timings are mapped back afterwards so your timestamps still match the file on your desk.
Two edge cases change the behaviour. If the detector finds no speech anywhere, FastScribe measures the file's mean volume. Under -50 dB it treats the file as near-silent and ships an empty transcript, the honest answer to a file with nothing in it, chosen over a page of confident sentences drawn from nowhere.
If the file is clearly audible but the detector rejects all of it, which happens with singing over music and sometimes with very quiet speech, one pass runs without the gate rather than handing you nothing. That pass is the one place invented text still reaches the page, and nothing on the transcript marks it as ungated. So if you upload a music-heavy file and text comes back, treat all of it as suspect.
The gate removes one class of failure completely. It does not make the model right about audio it can hear: buried words, people talking over each other and unfamiliar names still come back wrong, and noisy audio is still the hardest ordinary case.
What to check before you upload
Three checks before you upload, none of them slow, and between them they remove most of what produces invented text in the first place.
Trim the dead air at both ends. It is the single highest-value thing you can do to a file before uploading it, for the reason in the section above.
Measure the file rather than guessing. The free audio quality checker decodes your recording in the browser, nothing is uploaded, and tells you whether it is clipped, too quiet, or too close to its own background noise. Clipping matters most, because it cannot be repaired afterwards: the information is no longer in the file.
Make sure the recording opens with the speech. FastScribe detects the language from the opening of the file, so a recording that starts with music, room noise or a different language can be detected wrong, and a wrong detection does not produce a few odd words. It produces a whole transcript of confident nonsense. Getting a clean transcript from a rough recording covers the cleanup order that helps and the processing that makes things worse.
When an automated transcript is not enough
The line is not how often invention happens, because on clean speech it is rare. The line is whether catching it late still undoes the harm, and in some settings it does not.
Clinical notes and anything filed with a court are the clear cases. Reporting in 2024 found automated transcription inventing content in medical settings, and what made it serious was not a rate: one fabricated sentence in a patient record is a different kind of failure from a typo in a meeting summary. Use a person there, or check every line against the audio yourself.
Three other places this is the wrong tool. Nothing here runs while you talk: you upload a finished recording and get a transcript back, so if you need text during a meeting, look elsewhere. It does not write summaries or notes; you get the words that were said, and the shaping is yours. And it will not give you lyrics from a song, for the reason above.
Key takeaways
- A speech model cannot return nothing, so silence is the input most likely to produce invented text, and near-silence is worse than an empty file.
- Invented text reads exactly like correct text. You find it by checking where it collects, not by reading more carefully.
- Check the first and last thirty seconds of every transcript: that is where the polite-phrase shape lands.
- A block of identical repeated lines is the easiest shape to spot, visible in the shape of the page without reading a word.
- Trim dead air before uploading. It removes a whole class of error and buys back minutes against a length cap.
- The fourth shape, a confident wrong word in buried audio, survives every gate. Anything you intend to quote gets checked against the recording by a person.
Where FastScribe fits
FastScribe runs a voice activity detector before transcription, so the model is only ever shown the stretches that hold speech and non-speech audio never reaches the decoder. If the detector finds no speech and the file measures under -50 dB of mean volume, an empty transcript ships rather than a page of confident sentences drawn from nowhere. One case still runs ungated: an audible file the detector rejects wholesale, which in practice means singing over music, gets a single pass without the gate instead of returning nothing, and that pass is not marked on the transcript page, so treat a transcript of a music-heavy file as suspect throughout. None of this makes the model right about audio it can hear. Everything runs on our own hardware and no third-party AI service receives your audio, which is deleted the moment the transcript is ready. Your first file needs no account, up to 10 minutes and 50 MB, one per rolling week; a free account gives 5 files a day at up to 30 minutes and 100 MB each with transcripts kept for 7 days; Pro is $12 a month for files up to 5 hours and 2 GB with no daily cap. The transcript text stays in your history, which also means it sits on our servers and can be read there, so material you would not want a second person to read does not belong on any hosted service, this one included.
Frequently asked questions
Why does my transcript say "Thank you" when nobody said it?
Because the recording has a stretch with no speech in it, and the model filled the gap with the most ordinary phrase it knows. Models trained on internet video have heard "Thank you" and "Thanks for watching" at the end of countless files. FastScribe's voice activity detector exists to stop this, by never showing the model the silent stretch. If you see it anyway, check whether the file has a music bed, which is where the gate steps aside.
Why is the same sentence repeated over and over in my transcript?
That is the loop shape of transcription hallucination: the model sticks over sustained noise and emits the same line until the audio changes. It is the easiest form to spot, because you can see it in the shape of the page without reading a word. Delete the block, and if it sits over a passage that mattered, listen back to that timestamp rather than assuming nothing was said.
Does FastScribe hallucinate?
It runs the same class of speech model as everything else in this category, so yes, it can. What it does about it is run a voice activity detector before transcription, so the model never sees non-speech audio. Where the detector rejects an audible file wholesale, singing over music being the usual cause, one ungated pass runs instead of returning nothing, and that pass is not marked on the page.
Why did my transcript come back empty, or in the wrong language?
Empty means the file measured as near-silent, under -50 dB of mean volume, and the detector found no speech in it. An empty transcript is the deliberate answer there: the alternative is a page of invented sentences, which looks more useful and is worth less than nothing. A transcript in the wrong language usually means the file opens with music, room noise or another voice, because the language is detected from the opening. Trim to the speech and upload again.
Can better audio stop transcription hallucination completely?
It stops most of it. Trimming silence, recording close to the speaker and keeping music out remove the shapes that come from empty or non-speech audio, which is the majority. What survives is the fourth shape, a plausible wrong word where the speech was real but buried, and only a person checking the text against the recording catches that.
Try FastScribe on your own recording
One free transcription a week, no signup. Audio is deleted the moment your transcript is ready.
or drop it here · audio or video · one free transcription a week, no signup