You wrote "Tutankhamun". The narration says it correctly. The caption on screen reads "tootin car moon". If you make content in any niche with proper nouns — history, science, medicine, geography — you have seen this constantly.
The cause is architectural, not a bug
Almost every captioning tool follows the same sequence: generate the audio, then run speech recognition over that audio to produce captions. That second step throws away information you already had.
Speech recognition models are built to guess. Given audio that does not match a familiar word, they produce the most statistically plausible sequence of familiar words that sounds similar. For everyday speech this is exactly right. For "Tutankhamun", "Schrödinger" or "myocardial", it produces confident nonsense.
The frustrating part: the tool already had the correct spelling. Your script was the input to the voice synthesis. It was simply discarded before captioning.
Why the usual workarounds do not work
- Custom vocabulary lists — help marginally, need maintaining per video, and still fail on anything unlisted.
- Manual correction — works, but costs minutes per video and scales linearly with output.
- A better recognition model — reduces the error rate without eliminating it, because guessing is inherent to the approach.
Caption snapping
The fix is to use speech recognition for the one thing it is genuinely excellent at — timing — and take the words from where they were always correct: your script.
The process:
- Synthesise narration from your approved script.
- Run Whisper over that audio with word-level timestamps.
- Align the recognised sequence against the original script.
- Keep the timings; replace the words with the script's words.
Whisper hearing "tootin car moon" is now harmless. Those three tokens align back to "Tutankhamun", inherit that time range, and the caption displays your spelling with frame-accurate timing.
Why this matters beyond typos
Captions are read as authorship. A video that misspells the subject's name reads as careless regardless of how well researched it is — and in niches built on credibility, that is a real cost. It is also the kind of error viewers reliably point out in comments.
AutoShortsX snaps every caption to the approved script by default; there is nothing to configure. Download it or read about story and fact channels, where this matters most.