Auto captions and accent accuracy

Auto captions miss names, numbers, and accents first. There is no honest single word-error rate to put on a sales deck. The fix is read-back of the strings that can wreck the clip, then a new burn.

Automatic speech recognition is a listen, not a copy-edit. Versely's caption path, add captions to a video, transcribes spoken audio in 165 language codes and times the result. Accuracy depends on audio clarity; heavy accents, overlapping speakers, and background noise degrade the transcript. That limit is in the task, not a footnote. Treating the first render as final is how a guest named Siobhán becomes "Shivon" on a mute feed, which is the whole message.

This page is what goes wrong and how to catch it. It is not a leaderboard. Vendor WER numbers are measured on their test sets, with their normalisation, on their microphones. They do not travel to your podcast table.

What ASR actually hears

A recogniser maps acoustic frames to likely word sequences in the language you named. It is very good at common words in clean, majority-accented speech in that language. It is guessing otherwise.

The guesses cluster:

Proper nouns. Surnames, brand names, product SKUs, city names the model has rarely seen in this phonetic shape. "Versely" becomes a real English word that sounds close. A drug name becomes a food. The WER on that one token can be 100% on a clip whose overall rate looks fine because the other 90 words were "and" and "the."

Numbers. Prices, dosages, dates, "fifteen" versus "50," "point" versus a decimal. Some pipelines spell numbers as words; some emit digits. A mismatch against your on-screen graphic is a trust problem, not a style problem.

Negation and tiny words. "Can" / "can't," "now" / "not." WER treats them as one substitution. The ad treats them as the opposite claim.

Homophones. "There" / "their," "week" / "weak." Often a real English word, so a skim looks clean.

Code-switching. A Spanish product name in an English sentence, a Hindi aside, a French guest on an English host. One language code on a mixed take is a request to mishear the other language. The 165 codes include regional variants (es-MX vs es-ES, en-GB vs en-US); they are still one-language listens.

Crosstalk and rooms. Two hosts, a laugh, a café. The model will pick a speaker or invent a blend. Speaker IDs are an SDH job. Auto captions will not add them.

Accent is not a separate bug from these. It is a shift in the acoustics the language model was trained to expect. A recogniser tuned on broadcast American English will drop more on a Glaswegian take, a Lagos take, a Singapore take, or a second-language speaker with a strong L1. That is not a moral ranking of accents. It is where the training mass sat. You will not fix it by picking a "punchier" caption family. Snap on a wrong word is a wrong word that pops.

Names, numbers, code-switching

Build the QC list before you watch.

Write down, for this clip: every person named, every brand, every number the viewer is supposed to believe, every "not," every mixed-language token. Those are the read-back strings. If they are right, the rest of the transcript can be imperfect and the clip still ships. If they are wrong, a 2% overall error rate is irrelevant.

Set the language code to the speech that is actually on the file. A Nigerian English take is still English; try en-GB or en-US rather than a language the speaker is not using. A mostly-Spanish take with English product names wants es-MX (or the matching regional code) plus a manual pass on the English tokens. Do not set en-US "because the brand is American" on a Spanish voiceover.

If two languages are both first-class — a bilingual host — you are outside a single 165-code pass. Split the clip, caption each range with its code, or author the mixed lines. Hoping the model code-switches cleanly is how half the joke lands in the wrong language.

Numbers next to a graphic have to match the graphic. If the overlay says $49 and the caption says "forty-nine ninety-nine," you have two prices. Pick one in the transcript before you burn.

Names you care about should be in a glossary you actually look at. ASR will not learn your roster from the brand kit. Read them.

Accents without a fake WER

The FCC's caption quality rule for televised programming, 47 CFR § 79.1(j)(2), is qualitative: captions must be accurate, synchronous, complete, and placed. The 2014 order that adopted those standards (FCC 14-12) described them as non-quantitative. There is no 99% in the rule. Vendor pages that say "the FCC requires 99%" are collapsing an internal target onto a regulation that refused to pick one.

WCAG 2.2 Success Criterion 1.2.2 (Level A) is also non-numeric. Captions for prerecorded synchronised media must include dialogue, speaker identity, and meaningful non-speech. A WER of zero with no speaker labels can still fail that criterion on a two-hander.

So we will not put a Versely WER on this page, and we will not invent one for "Indian English" or "Scottish English." Those numbers would be a test-set cosplay. What you can say in a client email is: we set the matching language code, we read back names and numbers, we recaptioned the files that failed that list, and we watched the burn on mute.

If a particular speaker always fails — a thick accent on a cheap mic in a kitchen — the fix is the recording, not a different preset. Closer mic, less room, one person talking. Recaptioning the same muddy file in Reveal will not hear the surname better. It will animate the mistake.

Heavy accent plus music bed is two degradations. Drop or duck the bed for the caption pass if you can; restore it after. The captioner should hear speech.

Read-back is the QC

The workflow is short on purpose.

Transcribe and burn a preview of the first seconds if you are still choosing a look. Looks do not fix words.
Transcribe the full clip in the right language code, on a basic (1×) family until the words are right. Dynamic 2× is a look tax; do not pay it on a draft transcript.
Read the QC list out loud against the file. Pause on every name and number. Play the audio if you are not a native listener of that accent — the point of captions is that a mute viewer cannot.
Correct the wrong strings. Recaption that file.
Watch once on mute. If a line is unreadable, that is timing or contrast, not accent. Different pass.

Burned-in captions freeze the words into pixels. A later "we'll fix the name in the SRT" does not change the TikTok. Read before you burn.

The AI caption generator is the same engine. 165 codes, same "spot-check names, numbers and jargon" limit. Use it. Do not skip the read because the first thirty seconds looked fluent. Fluency is common words. The accent errors hide in the words you came to say.

A caption that is easy to read and wrong is worse than a slightly ugly caption that is right. Accent accuracy is the rightness pass. Style is later.

Read the next cut in Versely.

FAQ

Is there a captioner that "supports" my accent?

There is a language code, a clean mic, and a read-back. Marketing pages that promise a WER by dialect are measuring their corpus, not your room. Test on your speaker.

Should I pick a simpler caption style for heavy accents?

Style does not hear. Use a readable static family so a remaining error is at least checkable. Karaoke on a misheard name is how the wrong name becomes the hook.

Can I burn first and fix in a sidecar?

Not for short-form. The feed plays pixels. Fix the transcript, then burn again.