An Instagram transcript can contain words a speaker never said: research on Whisper has documented fluent hallucinated phrases, including text produced for audio with little or no speech. A readable result is not proof that the words are in the recording; play the source before quoting it. This article reports no first-party measurements because the original audio files and complete raw records needed for independent verification are not available. The discussion below is limited to published research and the benchmark-page copy reviewed here; it explains what those sources establish and how to check a transcript before quoting it.
What published evidence shows
There are different outcomes to distinguish: a system may misrecognize audible speech, label non-speech sound, or produce words without a spoken source. The research cited below documents the last category in Whisper systems. Those results do not establish an error rate for this site's browser model or for any individual Instagram Reel.
Why it happens at all
Speech recognition uses both the sound and patterns learned from training data to predict words. When the sound is weak, ambiguous, or absent, a prediction may still appear; sometimes it is a short label or stray word, and research has documented cases where outputs go further. That is why reading the transcript for fluency is not enough: compare it with the recording.
This failure mode has been studied in Whisper, the model family this site's recognizer belongs to. A 2026 arXiv paper describes fluent hallucinated transcripts on inputs containing little or no speech. Separately, Cornell researchers reported in 2024 that roughly 1% of the Whisper transcriptions in their study contained entire hallucinated phrases; their study analyzed more than 13,000 clips, including recordings from people with aphasia. Those findings concern the tested Whisper systems and datasets, not this site's particular browser model or your individual Reel. A Whisper repository discussion also records a user's report of repeated text after long non-speech intervals; that is an issue report, not a controlled measurement.
Sources
- Abbasihafshejani & Jadliwala, “Reducing Hallucinated Transcripts in Whisper via Hallucination Space Projection” (2026)
- Cornell Chronicle, “AI speech-to-text can hallucinate violent language” (June 11, 2024)
- OpenAI Whisper repository, PR #1253: user report of repeated text after long non-speech intervals (April 18, 2023)
What the three benchmark pages emphasize
The three sites selected for this comparison focus mainly on generating or editing transcripts. On the page copy we reviewed, they foreground accuracy, workflow, or caption features; that is a different emphasis from checking whether the output matches the original sound.
| Benchmark | What the reviewed page says | What we could verify about no-speech output |
|---|---|---|
| GetTheScript | Its accessible homepage says “95%+ accuracy on clear audio” and promotes link transcription, language selection, and subtitle exports. | The retrieved homepage copy did not describe a model returning invented spoken words for silence. |
| SpeakApp | Its homepage says “99% accuracy” and highlights recording, importing recordings, summaries, and rewriting. | The retrieved homepage copy did not discuss no-speech hallucinations. |
| Wayin.ai | The caption page describes animated captions and includes an unanswered accuracy question in the page text we retrieved. | The retrieved page text did not describe what its model outputs when audio has no speech. |
These notes describe only the page copy retrieved on November 4, 2026; they are not a claim about every feature or support page on those services. Their pages help users get or style text. This article adds a separate step: listen to the source before trusting or quoting any generated words.
Checks to make before you quote a transcript
Start with the source audio. These checks can flag a mismatch, but they do not replace listening to the words you plan to quote.
- Play the clip you transcribed. Listen to the source audio itself; do not rely on the transcript as proof of what was said.
- Look for repetition. A repeated word or phrase during a long quiet stretch is a warning sign, not proof by itself. Replay that part of the clip.
- Use duration as a rough check only. A transcript that seems unusually long or short is a reason to replay the recording, not a word-count test that proves accuracy.
- Interpret non-speech labels literally. A bracketed label may describe sound rather than speech. Do not quote it as dialogue, and do not assume every tool uses the same labels.
- Verify names and numbers against the recording. These details are consequential when quoted; the transcript alone cannot confirm them.
What to do when the transcript does not match
- Do not paste unverified lines into a report. A fabricated sentence can be harder to notice than a blank result, especially if it reads smoothly.
- Return to the original recording if possible. Avoid relying on a version that has been repeatedly converted or compressed, but remember that a cleaner file does not guarantee a correct transcript.
- Trim long silent or music-only sections and try again. This can make the speech easier to review; it does not guarantee that the model will stop guessing.
- Treat a blank result as inconclusive, not proof of a broken file. Check that the audio plays and contains speech. If a nonblank result appears for silence, do not pass it off as dialogue.
What this tool does and does not do
This site uses whisper-tiny.en, whose model weights are English-only. It runs in the browser and returns one plain block of text, with no timestamps, speaker labels, summary, or cleanup. This page does not present any first-party transcript measurements; research findings and competitor claims are attributed to their linked sources or the reviewed pages.
We publish no single accuracy percentage: a meaningful figure depends on the test audio, language, recording conditions, and scoring method. A result measured on one set of clips cannot promise the same performance on your Reel. This page does not report a first-party benchmark score; the numerical research findings above apply only to their cited studies.
FAQ
Can an Instagram transcript contain words nobody said?
It can. Cornell researchers documented Whisper outputs containing phrases that were not in the source audio, and a later paper studies hallucinated transcripts from inputs with little or no speech. These studies show a failure mode in tested systems; they do not predict whether a specific Reel or transcript tool will produce it.
Does a blank result mean nothing was wrong?
A blank result means this run returned no text; it does not tell you why. Check that the file plays and contains audible speech. If the tool instead returns words for a silent clip, do not treat them as dialogue.
Why does the model write anything at all when there is no speech?
Some speech-recognition systems can still return a prediction when the audio provides little evidence of speech. Learned language patterns may influence that prediction, but the output can be anything from a stray word to a longer passage; it is not necessarily fluent or consistent.
Does a smooth, readable transcript mean it is accurate?
No. Readability alone cannot establish that the words match the recording. Replay the relevant segment, especially before quoting names, figures, or full sentences.
Which of the popular transcript tools disclose this behaviour?
In the specific homepage and caption-page copy we reviewed for GetTheScript, SpeakApp, and Wayin.ai, we did not find an explanation of what the product might return for audio with no speech. This is limited to those retrieved pages, not a claim about every page or feature on those services.