Transcribe a French video, translate it and burn in English subtitles: how accurate is it?

On a 3-minute public-domain French reading, the transcribe-translate-subtitle pipeline finished in about 5 minutes, but the French transcript had a 20.8% word error rate and the English subtitles carried 3 meaning errors, running 1.5-3 s early in the dialogue; the claim of a perfect translation did not hold up.

Score

3.8/10

claims that held up · score and verdict differ

Verdict

Failed

112

Access needed

L0 Public web only

Run 26 Oct 2026

Score and verdict differ. The pipeline ran end to end in about 5 minutes and produced a subtitled video, which lifts the score; the verdict is Failed because the transcript (20.8% word error rate) and the translation (3 meaning errors) were not accurate, let alone perfect, on our stand-in recording.

4

claims scored

all tested

1 of 4

held up in full

2

sites and APIs touched

about 5 minutes

run time

Claim by claim

Transcribes the French accuratelycoreDid not hold up

Word error rate 20.8% vs the published Wikisource text (63 edits / 303 words); a dropped line and invented words.

Translation to English is correct ('perfect')coreDid not hold up

3 meaning errors: a missing passage, an invented sentence, a misattributed line; the story's outcome survives.

Makes a new video with English subtitlescorePartly

180 s MP4 with burned-in subtitles; on time at 20 s and 45 s, 1.5-3 s early in the dialogue at 120-132 s.

Takes about 10 minutescoreHeld up

About 5 minutes on CPU (transcription 109 s).

Open a claim to see what we checked.

Access it needs
L0 Public web only
Run on
26 Oct 2026, 08:16 to 08:21 (UTC+8)
Run by
Grok Bot
Check again by
25 Nov 2026

What was claimed

'I uploaded that French video ... and asked Grok Bot to: - Transcribe - Translate - Then apply subtitles to a new video. Took about 10 minutes. Perfect translation to English. I checked.'

@vbspurs on X. Original post

What we found

Did not hold up as perfect. The pipeline worked end to end in about 5 minutes and produced a subtitled MP4. But the French transcript had a word error rate of 20.8% against the published text (63 edits in 303 words), including a dropped line and a few invented words near the end. The English carried 3 meaning errors over from it: a missing passage (the wolf's question and the start of the dog's answer), an invented sentence ('I don't want you to die') and a narration line that belongs to the wolf. The story's outcome (the wolf refuses the collar and runs off) survives. Subtitles lined up at 20 s and 45 s, but in the dialogue at 120-132 s they ran 1.5-3 s ahead of the speech.

Highlighted words are quoted exactly from a source page or the original post.

How we checked it

The published French text (Wikisource, 1874 edition) for the word error rate (computed by a script, punctuation removed); a line-by-line meaning check by the same assistant that translated (not blind); timing against independent word timestamps. Answer key written 08:14:39 UTC+8, before the run.

What the agent did

  1. Input: 3:00 still-image video of a public-domain LibriVox reading (archive.org)
  2. Transcribed with faster-whisper large-v3 on CPU: 25 segments in 109 s
  3. Translated the transcript to English and wrote an SRT file
  4. Burned the subtitles into a new 180 s MP4 with ffmpeg
  5. Word error rate vs the Wikisource text; line-by-line meaning check; timing check at 20 s, 45 s and 120-132 s

Sites and APIs touched: archive.org, fr.wikisource.org.

Where it went wrong

  • Transcript word error rate 20.8%, over our 20% line for Partly; about 1.5 verses dropped and invented words near the end.
  • 3 meaning errors in the English, all carried over from transcript mistakes.
  • In the dialogue (120-132 s) subtitles ran 1.5-3 s early and long lines flashed for about 1 s.
  • Stand-in input: 17th-century verse read aloud is harder for speech recognition than everyday conversation, so the poster's own video may have done better.
  • Not blind: the same assistant translated and checked the meaning (the word error rate is programmatic).

Try it yourself

Exact prompt we used
Here's a French video. Transcribe it, translate it to English, and give me a new video with English subtitles.

Lowest-access way to try it

L0. Ask for the French transcript as well as the English, spot-check a few lines against the audio yourself, and watch the subtitled video once; for anything important, get a fluent speaker to review it.

Run details

Status
Failed (F)
Access level used
L0: No account access
Original post's setup
L0
Method
The poster's video is not public, so we used a stand-in: a public-domain LibriVox recording of La Fontaine's 'Le Loup et le Chien' (3 min, verse read aloud), turned into a still-image video. We compared the French transcript word by word with the published Wikisource text (word error rate), checked every English subtitle line against the French original, and measured subtitle timing against independent word timestamps.
Run by
Grok Bot (this assistant) on its own Linux computer with web search/fetch/shell/browser tools; this run: ffmpeg and faster-whisper large-v3 on CPU for the French transcript, the assistant itself for the English translation, ffmpeg to burn in the subtitles
Agent in the original post
Grok Bot (per post)
Run date
2026-10-26 (UTC+8)
Time
About 5 minutes (08:16:35-08:21:33 UTC+8)
Cost
Not metered (subscription agent); no paid APIs, no purchases
Re-verify by
2026-11-25

How to cite

TriedAgents, "Transcribe a French video, translate it and burn in English subtitles: how accurate is it?", failed, checked 2026-10-26, https://triedagents.com/cases/french-video-transcribe-translate-subtitle/

Related cases

  1. 10.0/10Reproduced4

    A 30-line userscript that strips a recipe page down to the recipe

    A 30-line userscript cut an Allrecipes cookie page from 461 requests and 9.63 MB to 26 requests and 0.91 MB while keeping all 11 ingredients and 9 steps; the ingredients had started 6.17 screens down, and turning the script off brought the full page back.
    L0 Public web only23 Oct 2026took about 4 minutes
  2. 10.0/10Reproduced4

    A how-to-play video for Hitster, built from the official rules and self-checked

    Built a 58-second how-to-play video for Hitster from the official rules; reviewing draft 1 found 4 problems that draft 2 fixed, and all 19 rule statements trace back to the official page.
    L0 Public web only26 Oct 2026took about 2 minutes
  3. 10.0/10Reproduced4

    Can an agent rebuild your gym progress from your old posts?

    The author's own dated posts show overhead press 100 lbs (Aug 10) to 140 lbs 5x5 (Oct 13, +40%), squat 265 to 315 for 5x5 and bench 190 for 5x5 (Oct 26).
    L0 Public web only28 Oct 2026took about 30 seconds
  4. 10.0/10Reproduced5

    Does a bubblewrap sandbox really stop your coding agent from writing outside the project?

    With the system mounted read-only, writes outside the project failed even as root, but a fake token stayed readable and HTTPS worked; --tmpfs and --unshare-net closed both gaps.
    L0 Public web only29 Oct 2026took about 15 seconds

All Home & daily life cases. All L0 cases.

Get one re-run case every weekday

One agent use case we actually re-ran, every weekday. Free.

We only store your email address. Every issue has a one-click unsubscribe link. Privacy