The Caption Keeps Talking After the Audio Stops
Automatic captions sometimes write whole sentences nobody spoke. Whisper’s own model card says why: predicting the next word and transcribing audio were trained as one job.
Put a microphone in front of someone who is thinking, and the audio goes quiet. Not silent — a room has a floor of noise — but empty of words. A few seconds pass. Then the caption box fills anyway.
Ask people what goes wrong with automatic captions and they describe one family of failure: a name spelled wrong, a technical term swapped for a homophone, an accent the model cannot follow. All of those are errors of hearing, and they share a fix — a better acoustic model, a cleaner signal, more context.
There is another class of error that shares none of it. The audio contains no such words at all, and the caption writes them anyway. This is not mishearing. It is speaking without having heard.
A sentence nobody said
The paper is Careless Whisper, published at FAccT in 2024. Its authors ran 13,140 short audio segments through Whisper and checked every output by hand.
One segment says, and only says, “pick the bread and peanut butter”. The transcript comes back as: “Take the bread and add butter. In a large mixing bowl, combine the softened butter.” That is a complete instruction, in the right register, for a sentence nobody in the room had spoken.
Another segment retells a scene from Snow White. The transcript ends with “For more information, visit www.FEMA.gov”. Elsewhere the model appends “Thank you for watching!” — the sign-off of a video that does not exist.
Look at what these fabrications have in common. They are not random noise. They have the shape of a transcript: a URL, a credit, a closing line. The model is not reaching for words. It is completing a kind of text it recognises.
| Error | In the audio | Fixed by |
|---|---|---|
| A word misheard | The sound | A better model |
| A language misjudged | The speech | More context |
| A line invented | Nothing | The pipeline |
The third row is the odd one, and it is why the rest of this article is about plumbing rather than about model sizes. No amount of acoustic accuracy addresses a sentence that had no sound to be accurate about.
The model card says it plainly
You do not have to take this from critics. OpenAI wrote the mechanism into Whisper’s model card, under limitations: “the predictions may include texts that are not actually spoken in the audio input (i.e. hallucination). We hypothesize that this happens because, given their general knowledge of language, the models combine trying to predict the next word in audio with trying to transcribe the audio itself.”
Read that twice, because it describes two jobs wearing one hat. Transcribing is listening. Predicting the next word is language modelling — the same machinery that finishes your sentence when you trail off mid-thought. Most of the time the two agree, because the next word is the word being spoken.
When the audio stops carrying evidence, the second job does not stop. It keeps predicting, and what it predicts is the most plausible continuation of a transcript. That is why the invented lines arrive with a URL attached: a URL is how a certain kind of transcript ends.
The training data explains the vocabulary without excusing it. The model card gives the figure — 680,000 hours of audio and matching transcripts collected from the internet. A model trained on that learns not only how words sound but what a transcript looks like, including how one tends to finish.
Silence is the variable
The sample is worth stating precisely, because it is the reason the finding holds up. 437 speakers, 13,140 segments, split almost evenly between people with aphasia and a control group. Aphasia is a language disorder that often leaves speech slow and full of pauses.
1.4% of segments produced a hallucination — a whole phrase or sentence with no counterpart in the audio. By group: 1.7% for aphasic speakers, 1.2% for controls.
Then the number that carries the argument. Segments that produced hallucinations had a higher share of non-vocal time: 29% of their duration against 26% for the rest. In seconds, 5.3 against 3.5.
The paper’s phrasing is careful — hallucinations “disproportionately occur” for speakers with longer non-vocal durations. That is a correlation, and three percentage points is not a law. But the direction survives every cut of the data, and it points somewhere uncomfortable: the people who pause most are the people a caption is most supposed to help.
It is not a stable bug
187 segments reliably produced hallucinations in the spring of 2023. The authors re-ran the same clips in December. 12 still did — 9 of 92 aphasic clips and 3 of 95 control clips.
That number cuts both ways. The good reading: most of it was fixed within seven months. The harder reading: this is not a defect that gets patched once and stays patched. It is a probability that moves with the model. A test that finds nothing today is not evidence of a fix, and a test that finds something today is not evidence of a permanent flaw.
Every chunk has to say something
An automatic caption is not one request. The audio is cut into chunks — around a third of a second each — and sent as it arrives, so that text can appear while the person is still speaking.
Every chunk must return a token sequence. There is no empty string in that vocabulary unless the pipeline builds one. For the model to say nothing, something upstream has to decide, explicitly, that this chunk contains no speech.
In most pipelines, voice activity detection does a different job. It finds the end of a sentence: the speaker stopped, so the line can be finalised and put on screen. That is a decision about when to display, not about whether to send.
The gap between those two is where the invented sentences live. A chunk of room tone is still a chunk. It still gets sent, still gets decoded, and the decoder still has to produce its most likely continuation.
This belongs to the architecture rather than to one vendor. The paper tested other providers on the same 187 clips and reported no comparable hallucinations from Google’s products — a claim that covers those clips and nothing else, so it should not be read as “everyone else is clean”. What it does establish is that the behaviour comes out of a choice of model and pipeline, not out of speech recognition itself.
Tablingo cuts audio into chunks of about a third of a second and sends each one as it arrives, which is why this article is not about somebody else’s product. The question it raises applies to every tool that streams audio into a decoder: does the pipeline have a way to report silence?
How to check it yourself
Two tests. The first: play ten to thirty seconds of music or room tone with automatic captions on, and see whether anything appears. It may not — the behaviour is probabilistic, as the December re-run showed. But when something does appear, it will have nothing to do with the audio: a URL, a credit, a sign-off.
The second: find a recording with long pauses — an older speaker, someone with a speech condition, a language learner — and read the transcript against what you hear. That is closer to the paper’s conditions, and it is where the failure is easiest to catch.
The score has no cell for silence
Neither test appears in the way captions are graded. The standard metric is word error rate: take the correct words, count the ones the system got wrong. It assumes a correct string exists.
When the audio contains no words, the correct output is nothing. Word error rate has no cell for that. A fabricated sentence costs a handful of word errors; a correct silence costs none. On that scoreboard the system that invents and the system that stays quiet are hard to tell apart — or worse, the one that invents looks busier.
So what is missing is not accuracy. It is an explicit empty output: a way for the system to report that this stretch of audio held no speech. A caption that can say “nothing was said here” is more trustworthy than one that always has something to say, even when its word error rate is worse.
That holds on both sides of the live and recorded line. Waiting for the stream to end does not remove the judgement — a full recording contains more silence, not less, and the decoder still needs somewhere to put it. Waiting buys a better look at the context. It does not buy a way to say nothing.
If you want to try ours, it’s called Tablingo.