They are speech recognition mistakes. YouTube's automatic captions are written by machine learning, in the video's default language only, and YouTube's own help page warns that they "might misrepresent the spoken content" because of accents, dialects or background noise. Over music, noise or silence, the recognizer still tries to write something, and a few short words come out again and again. Viewers have described "foreign" and "heat" appearing this way in online discussions, and "thank you" on silence is a documented habit of at least one widely used speech recognition model.
YouTube has not published an explanation for these particular words. This guide separates what YouTube documents from what users and researchers have observed, as of 19 October 2026. If the words appear on your own video, you can fix them in YouTube Studio in a few minutes.
What the strange words mean
| What you see | When it tends to appear | What is known | Source and date |
|---|---|---|---|
| [Music], [Applause], [Laughter] | Music, clapping, laughing | Deliberate sound labels in automatic captions | Google Research blog, 23 Mar 2017 |
| Words in all caps, sighs, gasps | Shouting, reactions | Expressive captions on English videos | YouTube blog, 2 Dec 2025 |
| foreign | Noise and music the recognizer could not place | User reports about older captions; YouTube has not explained it | Hacker News, 22 Jul 2025 |
| heat, often repeated | Silence and noise | User reports only; YouTube has not explained it | Hacker News, 22 Jul 2025 |
| thank you, thanks for watching | Silence, especially at the start or end | Documented in OpenAI's Whisper; not confirmed for YouTube | Cornell, 11 Jun 2024; OpenAI forum, 11 Mar 2023 |
The first two rows are features. The last three are errors.
Sound labels are deliberate
Some bracketed words are meant to be there. In a Google Research post of 23 March 2017, an engineer on Google's Sound Understanding team explained that automatic captions would start labeling sounds, and that "for the initial launch of this feature, we chose [APPLAUSE], [MUSIC] and [LAUGHTER]", because those are among the sounds people most often caption by hand.
Since then YouTube has added expressive captions. YouTube's automatic captioning help says English automatic captions show "the intensity of speech in all caps, expressions of sounds like sighs and gasps, as well as noises from the environment". A YouTube blog post of 2 December 2025 says these are on English videos across all devices, starting with videos uploaded after October 2025. So a caption in capitals, or a [gasp], is the system working as designed.
Why "heat" and "thank you" appear over music and silence
Speech recognition turns sound into the most likely words. It is good at clear speech and weaker where there is none. Music with vocals, a beat, crowd noise or a long pause all leave it guessing. Describing OpenAI's Whisper model, Cornell researcher Allison Koenecke put it this way: the technology seems to be "seeding silence as some sort of word".
"Heat" is the clearest example on YouTube. In a Hacker News discussion on 22 July 2025, one user described YouTube's newer captioning model writing "heat", "often repeated", in silence and noise it could not classify, where the older model had written "foreign". These are user reports, not statements from YouTube, but they match what viewers search for. If that user is right, it also explains why both words turn up: older captions say one, newer captions the other.
"Thank you" has better evidence, though not for YouTube itself. A Cornell study presented in June 2024 found that roughly 1% of transcriptions by OpenAI's Whisper contained entire made-up phrases, and that Whisper wrote "Thank you" for a silent audio file; the researchers think longer pauses make these errors more likely. Koenecke added that OpenAI has improved the model since the study was carried out. In March 2023, a user on OpenAI's developer forum reported that Whisper fills empty sections at the start or end with "thanks for watching". And in a GitHub discussion opened on 13 June 2025, users showed Whisper turning complete silence into a subtitle credit line in Arabic; one commenter suggested this is because such models learn from video audio paired with subtitles, where the quiet end of a video often carries a sign-off or a credit.
Whisper is OpenAI's model, not YouTube's. Google said in March 2023 that its own Universal Speech Model is "for use in YouTube (e.g., for closed captions)", and it has not said why its captions show these words. The pattern is the same, though: a quiet or musical stretch, a model that must output text, and a short, common phrase as its best guess. Music is the other big trigger, and how background music and noise affect automatic subtitles covers that side in detail.
Is "foreign" about other languages?
Possibly, though nobody has confirmed it. YouTube's help on adding subtitles states that "automatic captions will be in the video's default language only". Each video gets captions in one language. If the video is set to English and someone speaks Hindi, Spanish or Korean, the English recognizer cannot write those words, and YouTube lists "multiple languages at the same time" as one reason automatic captions fail.
The user reports we found are about noise rather than language. In the same Hacker News thread, one user recalled that YouTube's earlier captions marked noise and music they could not understand as "[foreign]": "If it couldn't understand it, it was 'foreign'". A second user remembered it showing as a plain word rather than in brackets. Speech in a language the track does not cover is also audio the recognizer cannot understand, so it is a likely place for "foreign" to appear, though neither YouTube nor those users say so directly.
For viewers, auto-translate does not help here: it translates the caption text, and the caption text is already wrong. If the creator added a caption track in the other language, pick it from the captions menu.
What you can do as a viewer
You cannot correct the words on someone else's video, but you can work around them:
- Turn the captions off. If the errors get in the way, click the Closed captions button again; YouTube's caption settings help describes it as an on and off switch.
- Choose a different track. If the creator uploaded their own captions, pick that track in the captions menu instead of the automatic one. A creator's own track is more likely to have been checked by a person.
- Get a second attempt. Chrome's Live Caption, under Settings and Accessibility, runs its own speech recognition on the audio playing in the browser. It can also make mistakes, but they will not be the same ones.
- Tell the creator. A comment with the timestamp of a wrong line is the quickest way to get it fixed for everyone.
Fixing your own video's captions
If these words are on your own upload, open YouTube Studio, go to Subtitles, select the video, and click More next to the automatic track to edit or remove the wrong lines. Check the opening, the ending, music sections and any lines in another language first, because that is where these errors cluster. YouTube auto captions missing or late walks through the Studio steps and the other caption problems YouTube lists.
On someone else's video you cannot edit the captions, since YouTube ended community contributions in 2020. Adding subtitles to someone else's YouTube video explains what still works.
Any speech recognition can mishear, so check before you publish
These errors are not unique to YouTube. Every automatic captioning tool, including CapsAI, uses speech recognition, and every one can mishear music, silence and noise. The difference that matters is whether you can see and fix the text before viewers do. Common mistakes in AI-generated subtitles lists the usual ones.
A short checklist for any caption draft:
- Read the first and last ten seconds. Silent intros and outros are where phantom phrases appear.
- Watch every music section. Delete words written over a beat or a song unless they are real lyrics you want shown.
- Check other-language lines. If the video mixes languages, make sure each line is in the language actually spoken.
- Fix names, numbers and brand terms. They are the most common real mishearings.
- Watch once with the sound off. If a caption makes no sense without audio, it is probably wrong.
With CapsAI you choose the spoken language before upload, from English, Hinglish, Hindi, Bengali, Telugu, Marathi, Tamil, Urdu, Gujarati, Kannada and Punjabi, then edit every word and timing in the browser before you export an MP4 with the captions burned in. Burned-in captions do not replace YouTube's track, so review that too, or turn it off if it duplicates what is on screen. The YouTube subtitle generator page covers this route for videos up to 10 minutes.
What to do next
If odd words are on your own video, fix the automatic track in YouTube Studio first; it is free and takes minutes. If you want styled captions you have checked line by line, open the uploader with English selected, or pick the language spoken. A new account has 3 free minutes to start, and packs are a one-time payment from $4 (₹49 in India, €4 in the euro area).
