Skip to the content

Case study

Why AI video drops the presenter on lines with prices and decimals

A client's video team found good presenter lines replaced by stock footage whenever the script held a decimal. The automatic check was reading one number as two, and the captions lost the decimal mark. We traced it and fixed it within a day.

Key takeaways

  • If the presenter vanishes only on lines with figures, suspect the automatic check's reading of numbers before you suspect the presenter.
  • A rejected take that sounds correct is evidence about the check, not about the recording.
  • Captions that drop a decimal mark can make a correct figure look careless.
  • A video tool should be judged on the lines that carry your claims, not only on plain sentences.
  • A fault that follows one kind of line usually has one cause, so look across recent videos for the pattern.

When an AI video drops the presenter on number lines, the cause is often the automatic check, not the person on camera. We traced that fault for a client and fixed it within a day; the fix went live on their platform the same day it was traced.

What the business saw: good lines coming out without the presenter

A client's video team makes long ads with an AI presenter who speaks the script line by line. In one long ad, a few lines came out with stock footage and the cloned voice instead of the presenter. Every one of those lines held a figure with a decimal comma, where a comma sits inside the number, such as a speed in seconds or a success rate. On the same lines the captions showed 04 instead of 0,4.

In a sales ad the figures are the proof: the speed, the success rate, the price. Those are the lines where the presenter matters most, and the lines that lost him.

Why AI video drops the presenter on number lines

The automatic check listens to each take and compares it with the script, so a mumbled or skipped word does not reach the ad. That check is only as good as the way it reads a number. Scripts for ads are full of prices, percentages and decimals, which are exactly the lines a check can misread. A strict check can be useful and still be wrong: it is reliable only where it reads the script the same way the script was written. A caption that shows the wrong figure can make a viewer doubt the claim it supports.

What our engineer found: two systems disagreeing about a word

The check's verdict on each failed take was one word missing and two words added, yet played back the presenter had said the line perfectly. The turning point was comparing the two things the transcription service, the software that turns the recording into text, returns for the same recording: a full text and a separate list of single words with timings. The full text wrote the number whole, as the script did. The word list split it at the comma into two numbers, nought and four. The check was built from the word list, so it heard two numbers where the script had one and refused a correct take. Both outputs came from the same recording, so the disagreement was inside the tool, not in the presenter's delivery.

Searching about six weeks of the platform's history showed this was no accident of one ad. Every presenter take of a line with a decimal had failed the same way, with commas and with points, prices included. A second fault sat in the captions of lines the cloned voice read: the voice service records a timing for each character, and grouping those characters back into words stripped every comma and full stop, including the one inside a number.

What we changed

Now the check puts a number back together. When the full text shows a number with a comma or point inside it, the separate pieces in the word list are joined into one word again, timed from the opening piece to the closing piece. The joining only happens where the pieces match that number in order, so two numbers spoken one after another are not joined by mistake. A percent sign after the figure stays part of the number. The same repair feeds the captions, so they show the number as written. In the voice step, punctuation is removed only at the edges of a word, never inside it. Each real line that had failed became a test the check must pass, alongside lines that must be left alone.

What the business has now

The fix is live on the client's platform. Presenter takes of lines with decimal numbers now pass the check when the presenter said them correctly, so those lines keep the presenter instead of falling back to stock footage. Captions show the figure as the script writes it. Ads made before keep what they were made with; a line made again now comes out right.

What you can check yourself

  • Look through your last few AI videos for lines with prices, percentages or decimals. Is the presenter on screen for them, and does the caption show the number exactly as written?
  • Ask whoever makes your videos which lines the automatic checks threw out, and whether any of them were actually fine.
  • Listen to one rejected take against its script. If it sounds correct, ask how the tool reads numbers: as one word or split into pieces.
  • Write one short test script using the number formats you really use, including a decimal comma, a decimal point, a price and a percentage. Watch whether the presenter stays and the caption keeps the mark.
  • Check a line the cloned voice reads as well as one the presenter reads. The two paths can fail in different ways.

Questions people ask

Why did the AI presenter disappear only on some lines?

The failure followed the punctuation inside a figure. A tool that treats a comma or point as a word break can count one number as two, so its comparison with the script fails even when the recording is clean. That is why the same presenter can sound perfect and still be rejected on those lines.

Why does my video caption show 04 instead of 0,4?

The caption step has probably removed punctuation from inside the number while rebuilding words from timed pieces. The mark is gone before the caption appears, not changed on screen. Ask your supplier whether punctuation is stripped only at the edges of words. If it is removed everywhere, decimal captions will keep losing the mark.

The check says words were missing but the presenter said the line correctly — what does that mean?

It means the check and the script are not counting the same units. A decimal can be one item in the script and two in the tool's word list, so the tool reports a missing word and extra words. The useful next step is to ask for the two texts it compared, not to re-record immediately. If the full text matches and only the word list splits the number, the take is likely fine.

Will my older videos be fixed too?

No, a fix to the check changes how a new take is judged, not the files already made. Old ads keep the presenter gaps and caption errors they were published with. If an old line still matters, make that line again after the fix and compare the new version with the old one.

How do I test a video tool before trusting it with a real ad?

Give the supplier a practice script and ask for the raw result before any editing. Compare three things: whether the presenter stayed on every line, whether the caption matches the script character for character, and what the automatic check said about each take. Then listen to any take the tool rejected. A tool that cannot explain a rejection is hard to trust with a live campaign.

From Tech Concierge

Want this handled for your business?

Your own engineer builds and runs the website, the app, the CRM and the AI assistant that answers your customers.

The first 7 days are free.

Read next

Case studyVideo and content

Why your AI video presenter sounds sped up

A calm explainer came back from the AI video tool with the presenter talking like a recording on fast-forward, and the longest lines read by someone else. The voice was fine; a pace setting and a clip length were not.

4 min read

Case studyVideo and content

Why the notes in your video script never reach the finished ad

A video tool that reads only the spoken words will drop the writer's notes on footage, on-screen text, sound and acted scenes. A client's team found the gap only after watching the finished ad; here is what we changed and what to check in your own videos.

5 min read

Case studyVideo and content

Why an AI image adds an extra person to your two-person scene

A client's video team wanted a two-person conversation ad. The AI image tool kept drawing the person each speaker was looking at, so the cut would show four people. Describing the framing, not who must not be there, fixed it.

4 min read

We count visits to this blog with our own analytics, never Google's, and share what you read with nobody. May we remember this browser with a cookie? See the privacy policy.