Why Your Transcript Says "Thank You" When Nobody Said It
AI transcripts sometimes contain "Thank you." or a subtitle credit nobody said. Why Whisper writes them over silence and music, the common ones by language, and when they matter.
Short answer
Lines like "Thank you.", "Music" or "Subtitles by" that nobody said are hallucinations. The speech model was given music, applause or silence and wrote the text it most often saw over such audio in its training data. They are not in the recording, and they sit where nobody is speaking: the start, the end, or long pauses.
You open an AI transcript and the first line is "Thank you." Then another "Thank you." Nobody in the video says it. Further down there may be a lone "Music", a "The End", or a credit like "Subtitles by the Amara.org community".
These are hallucinations, and they are among the best-known quirks of Whisper, the speech model behind most AI transcription tools, SubGrab included.
How Common They Are
Of the 6,217 English AI transcripts in SubGrab's history, 253 begin with a line that is only "Thank you.", 77 with "Music" and 56 with "The End". That is about 6% of them, and almost every one marks a stretch with no speech at the start of the recording: an intro jingle, a waiting screen, silence before a meeting starts.
Why the Model Writes Them
Whisper was trained on hundreds of thousands of hours of audio from the web, paired with the subtitles people had written for it. Subtitle files are full of text that is not speech: "Thank you for watching" over an outro, "[Music]" over a jingle, and the subtitler's own credit at the very end.
So when the model is handed audio with no speech in it, it does what it learned: it writes the text that most often accompanied that kind of audio. It is not hearing those words. It is predicting them.
The Common Ones, by Language
| Language | What appears | What it means |
|---|---|---|
| English | "Thank you.", "Thank you for watching.", "you", "Music", "The End" | Outro and music captions |
| English | "Transcribed by ESO, translated by" | A subtitler's credit |
| French | "Sous-titrage ST' 501", "Musique" | A subtitling credit, music |
| Russian | "Субтитры создавал DimaTorzok" | "Subtitles created by DimaTorzok" |
| Arabic | "ترجمة نانسي قنقر" | "Translation by Nancy Qanqar" |
| Chinese | "中文字幕志愿者" followed by a name | "Chinese subtitle volunteer" |
| Welsh | "Diolch yn fawr iawn am wylio'r fideo." | "Thank you very much for watching the video." |
We have seen every one of these in real transcripts. The Welsh one is a classic: when the model cannot judge the language of a silent stretch, it sometimes picks Welsh.
Where They Show Up
- At the start, over an intro, a countdown or silence.
- At the end, over outro music or a sign-off screen.
- In long pauses, over a music break or a dead stretch in a meeting.
A tell-tale sign: a line spanning exactly 30 seconds that holds one or two words. Real speech is rarely that sparse.
When It Matters: The Wrong-Language Problem
A phantom line at the start is usually harmless. On a non-English video it can be a warning. Whisper decides a recording's language from its first 30 seconds, and if those seconds hold only music, it tends to guess English and write the real speech out as an English translation.
SubGrab stopped letting the opening decide on 13 October 2026: the language is now decided from samples spread across the whole recording, and samples that only produce lines like these do not count. The full story is in why an AI transcript can come back in English.
How to Clean Them Up
Download the transcript as TXT or SRT and delete the phantom lines. In an SRT file, each one is its own numbered block; delete the block and the subtitles around it keep their timing.
SubGrab does not delete them automatically. A speaker can really say "thank you", and removing a line someone actually said is worse than leaving a harmless phantom one.
Frequently asked
Is the rest of my transcript wrong too?
Usually not. These lines appear where there is no speech, and the speech around them is transcribed normally. The exception is a hallucination at the very start of a non-English video: it can mean the language was judged from that silent opening, and the transcript may have come back in English. Since 13 October 2026 SubGrab no longer lets the opening decide the language.
Why does my French transcript say "Sous-titrage ST' 501"?
It is the same thing in French: a subtitle credit the model saw in its training data and writes over music or silence. Russian transcripts get "Субтитры создавал DimaTorzok", Arabic ones "ترجمة نانسي قنقر" and Chinese ones a subtitle-volunteer credit. None of them is in the recording.
Does SubGrab remove these lines automatically?
No. A speaker can really say "thank you", and deleting a line someone actually said is worse than leaving a harmless phantom one. They are easy to remove by hand: delete the lines from the downloaded TXT or SRT file.