AI Transcript Came Back in English Instead of the Video's Language? Here's Why
A French, Korean or Hindi video whose AI transcript reads like an English translation usually opens with music or silence. Why it happens, how often, and how SubGrab fixed it.
Short answer
When a video opens with music, a countdown or silence, speech recognition can misjudge its language. Whisper decides from the first 30 seconds, guesses English, and then writes the real speech out as an English translation. Since 13 October 2026, SubGrab decides the language from speech sampled across the whole video, so the opening no longer decides it.
You paste a link to a French conference, a Korean drama or a Hindi lecture, and the transcript comes back in English. Not garbled English: fluent, sensible English, with the odd line still in French. It reads like a translation because it is one.
This was the most damaging mistake an AI transcript could make, because nothing about it looks broken. Here is why it happened, how often, and what SubGrab does about it since 13 October 2026.
What It Looks Like
- The first lines are "Thank you." or "Music" over a stretch where nobody spoke.
- Most of the transcript is in English, but some lines are still in the original language.
- The transcript is labelled English.
A customer sent us a three-hour French policy hearing in October. In the first 45 minutes, about 60% of the lines were still French and the rest English. In the remaining two hours and fifteen minutes, almost every line was English.
Why It Happens
Whisper, the speech model behind most AI transcription tools including SubGrab, decides a recording's language once, from the first 30 seconds it is given. If those 30 seconds hold music, applause, a countdown or silence, there is no speech to judge. It then usually guesses English, often while "transcribing" the silence as "Thank you." (see why transcripts say "Thank you" when nobody did).
Having guessed, it transcribes everything after it in English, and for French speech that means translating it.
We measured this on that French hearing, using the same model and settings as production:
| Audio sent | Language detected | What came back |
|---|---|---|
| The first 10 minutes, starting with 53 s of room sound | English | "Thank you." twice, then English |
| The same audio, starting one minute in | French | French |
| A section from 45 minutes in | French | French |
| The same section, told the language is English | English | A clean English translation |
The result was the same every time we ran it. The opening alone decided.
Why Long Recordings Were Worse
Recordings over roughly two hours are split into 45-minute parts and transcribed in parallel. SubGrab used to let the first part's guess decide every other part, so one bad opening reached the whole recording.
Being told the wrong language is worse than guessing it wrong. In that hearing, the first part, where English was only guessed, kept 61.5% of its lines in French. The later parts, where English was set explicitly, kept between 2% and 9%.
Before the change, about 4% of the AI transcripts we had labelled English were really in another language: German, Korean, Chinese, Japanese, Persian, Urdu, Hindi, Thai, Vietnamese, Spanish, Arabic and French among them. The longer the recording, the more likely:
| Recording length | Transcripts affected |
|---|---|
| Under 10 minutes | 2.4% |
| 10 to 45 minutes | 4.7% |
| 45 minutes to about 2 hours | 7.6% |
| Over about 2 hours | 14.6% |
Long recordings are livestreams, webinars and events, and those tend to open with a waiting screen.
How SubGrab Decides the Language Now
Since 13 October 2026:
1. SubGrab takes four 30-second samples from the audio: the start, and 20%, 50% and 80% of the way through.
2. It detects the language of each sample separately. A sample that produces almost no text, which is what music, silence and "Thank you." look like, does not vote.
3. The language with the most votes is used for the whole recording, including every part of a long one.
4. If there is no clear winner, nothing is forced, and the recording is transcribed exactly as before.
We tested the rule on 25 real recordings from our own history before shipping it. It fixed all 10 non-English recordings that had come back in English, changed none of the 13 that were already right, and left both bilingual recordings alone. The samples add about a second to a job.
What It Cannot Fix
- Recordings in two languages. A video that is genuinely half one language and half another gets one language. For clean results in both, transcribe each section separately.
- Recordings that are mostly music or singing. If almost none of the samples land on speech, there is nothing to vote with, and the opening decides as before.
YouTube Captions in the Wrong Language Are a Different Problem
If a YouTube transcript comes back in a language the speaker never used, the video has probably been auto-dubbed and the captions belong to a dub. That has its own cause and fix: see YouTube transcripts in the wrong language.
For what to expect from each language, see how to transcribe a video in 90+ languages.
Already Have a Transcript in the Wrong Language?
Transcripts made before 13 October 2026 are not re-run automatically. Write to help@subgrab.com with the transcript's link and we will look at it.
Frequently asked
Does SubGrab translate videos into English?
No. SubGrab transcribes in the language that is spoken. An English transcript of a French or Korean video is a mistake in deciding the language, not a translation feature. Since 13 October 2026 the language is decided from four samples spread through the video instead of its first 30 seconds, which removes the most common cause.
Why does my transcript start with "Thank you." when nobody says it?
Because the opening had no speech in it, and the speech model wrote the phrase it most often saw over music and silence in its training data. It is a hallucination, not something in the recording. When it appears at the start, it is also a sign that the opening was used to judge the language.
What happens with a video that uses two languages?
If the samples split evenly between two languages, SubGrab forces neither, and each part of the audio is transcribed in the language it starts in, as before. The transcript then favours one language. For a clean result in both, transcribe each language's section as its own file.
I already have a transcript in the wrong language. What can I do?
Transcripts made before 13 October 2026 are not re-run automatically. Write to help@subgrab.com with the transcript's link and we will look at it.