ASR misses names and numbers
The pass that ships is a read of names, SKUs and prices. ASR is a draft. Burn after that pass. This is not a model page.
Automatic speech recognition misses names and numbers. The pass that actually ships is a human read of people, SKUs, and prices. Everything else — preset, karaoke, language code — is downstream of that list. This is not a model page. It is the edit.
The draft is fluent; the miss is the claim
ASR is good at common words in clean audio. That fluency is the trap. The first thirty seconds look like English, or Spanish, or Korean. The error is hiding in the word you came to say.
People. Guests, founders, talent, the surgeon in a clinic clip, the teacher in a webinar. Siobhán, Nguyen, a hyphenated brand used as a surname. Mute autoplay will present the phonetic neighbour as fact. Comments will correct you. Ads will not get a correction; they will get a complaint.
Places and orgs. Hospitals, universities, product lines named after towns. The recogniser has a prior. Your proper noun may not be it.
SKUs and model strings. “XJ-240,” “Pro 4K,” a lot number. Alphanumeric tokens are where the listen improvises. If the box is in the frame, the box is source of truth. If the box is not in the frame, the script is.
Prices and quantities. “Fifteen ninety-nine,” “fifty percent,” a dosage, a phone number, a date. Digit errors are business errors. A muted UGC ad with the wrong price is the wrong ad.
Legal and medical words. The qualifier the lawyer added, the ingredient, the statute. ASR will substitute a common neighbour. You will not notice unless you are looking for that string.
None of this is fixed by switching from BASIC to DYNAMIC. 2× animation does not spell. None of it is fixed by Snap. Karaoke on a wrong name is a highlighted wrong name. 165 language codes pick the listen (en-US, es-MX). They do not certify the SKU.
Overlap, music under the voice, and a Zoom mix make the miss denser. You still run the same list. You just start from a worse draft.
The pass is a list, not a watch-through
A casual playback is how you miss the one number. Make a list from the brief before you caption:
- Talent and guest names, spelled as they will be tagged
- Product names and SKUs, copied from the PDP
- Prices, discounts, durations
- Any word a regulator or a buyer will screenshot
Then transcribe. Then search the transcript for each item. Then fix the cue. Then burn.
Order matters. Burn-in is pixels. Patching a price after the Reel is live is a new render and a new upload. Sidecar SRT on YouTube long-form can be patched in Studio. Short-form cannot. The pass is cheaper than the recaption.
Do not hyphenate a name to win a line wrap while you are in this pass. Do not “fix” a number by writing it in words if the on-screen graphic shows digits, or the other way around, unless you meant to. Consistency with the plate is part of accuracy.
Speaker IDs are a related pass on multi-speaker clips. ASR will not reliably label a webinar. If two voices share the mix, you add PRIYA: where the face does not do the work. That is still this job: who said the number.
If the take has no speech, stop. There is no ASR miss because there was no listen. Write the overlay. Running the captioner to “see if it caught the name” on a silent product plate is a spend on empty audio.
Burn after the list is green
Only then pick a look.
- BASIC, 21 presets at 1×, two-line, sentence case, for ads, courses, anything that will be reviewed.
- DYNAMIC, 9 at 2×, only if the organic hook is a spoken phrase you can still finish with the corrected words.
Preview style on the first seconds. Read names on the whole file. Those are different QCs. Teams reverse them: they fall in love with Glass on a wrong SKU.
Language: set the code to what was spoken. A correct English SKU inside a Spanish sentence still needs a mixed-line decision. Do not expect one code to honour both.
Destinations:
- Reels / TikTok / Shorts: burn the corrected transcript. The feed will not play your SRT.
- YouTube long-form / course: keep the corrected SRT as a sidecar too. Convert with the VTT to SRT converter if you need the other extension.
- Paid: the price in the caption must match the price in the legal overlay and the landing page. Three surfaces, one number.
This is not a Veo page, not a Seedance page, not a “which ASR is best” bake-off. The engine on add captions to a video is the VEED listen. It is a draft. Your list is the product.
Make the pass boring enough to repeat
Write the list into the series bible so episode four does not rediscover the spelling of the guest from episode one. Pin a house caption style so QC is not also an art direction meeting. Estimate cost before a long DYNAMIC render so you are not paying 2× to typeset a draft you have not read.
The AI caption generator will give you the draft quickly. That speed is only useful if the read still happens. Add subtitles automatically is the same engine with a BASIC intent — still names, still numbers, still a pass.
Mute autoplay does not forgive a miss. The scroller does not hear you say the name correctly over a wrong caption. They see the wrong caption. That is the whole format.
ASR misses names and numbers. The ship file is the file after someone checked the list against the brief, the pack, and the landing page. Then you burn. A second pair of eyes on those three strings is cheaper than a paid comment thread explaining that the SKU was right in the audio.
Read the names, then burn in Versely.
FAQ
Can I skip the read if the audio is clean?
No. Clean audio improves common words. Names, SKUs, and prices are still the failure mode. Search the transcript for the brief’s list, then burn.
Does a better caption preset reduce name errors?
No. BASIC and DYNAMIC change drawing and the 1× / 2× multiplier. Word-highlight changes the clock. Spelling is the edit pass.
What if the box on screen and the spoken price disagree?
Pick the source of truth (usually the offer you will honour) and make type match it. Then fix the other surface — the mouth via a recut, or the graphic via an overlay — so the ad does not argue with itself.