What auto captions get wrong

Auto captions fail in patterns you can name. This is a field guide to seven of them: what they look like on the file, what they are not, and what to fix so you do not recaption the wrong layer. It is not the case for captions on a muted feed. It is the error catalog you walk after the generator returns.

Speech-to-text plus alignment plus layout are three machines. A "bad caption" complaint is usually one of them. Read the cue, then pick the machine.

Names, numbers, and homophones

These three are transcription misses. The mouth and the clock can be perfect. The string is still wrong. On a burned file the wrong string is the ad.

1. Coined names and brands. ASR emits a common word that sounds like your SKU. "Versely" becomes a real adverb. A serum name becomes two English words. You will not hear the error if you watch with sound; you will see it if you watch on mute, which is how the feed watches. Spot it by reading every proper noun in the transcript before burn-in. Keep a stoplist per client: brand, product, founder, city. If those four strings are right, most of the embarrassment is gone. Fix: correct the transcript, recaption. Do not turn the audio up.

2. Numbers, prices, dosages, SKUs. Digits are where models guess. "$14.99" becomes "fourteen ninety-nine" or "1499" or "forty." A batch code, a shade number, a timestamp in the script — same family. Spot it by searching the transcript for numerals and for number-words. Compare to the pack shot. If the on-screen bottle says 30 ml and the cue says 13, the caption is now a claim. Fix: type the canonical form you want on screen (digits or words, one style per campaign) and recaption. Do not let the engine pick.

3. Homophones that survive a mute skim. "Their / there," "week / weak," "in stock / instantly." These look like English. They pass a spellcheck. They fail a product check. Spot it by reading the cue as a claim, not as audio. If the line could be a different true sentence, it is a homophone risk. Brand slogans and legal phrases belong on the same stoplist as names. Fix: replace the word, recaption. A second listen with sound will not catch it, because the talker said the right one.

These three are why "are auto captions accurate enough to burn without reading?" is no. Accuracy on ordinary dialogue can be fine. Accuracy on the one string the ad depends on is the test, and that string is usually a name, a number, or a slogan.

Crosstalk, music-as-words, and the two timing bugs

These four are identification and clock misses. The words may be right and the file still unusable.

4. Crosstalk concatenated into one speaker. Two mouths, one cue. The engine interleaves words into a sentence nobody said. Interviews, a founder and a customer, a podcast clip, a store walk-through with a second mic — this is the default, not an edge case. Spot it: a cue that is grammatically broken, or that switches person mid-line without a label. Fix: split the lines, label speakers if the player (or the burn) cannot place cues under each face, recaption. Auto captions will not emit (Maya) unless you add it. If you need speaker IDs and [music] cues, that is an SDH pass on a sidecar, not a hope about ASR.

5. Music and SFX ingested as vocabulary. A bed, a whoosh, a laugh, a notification chime — the model emits a word that was never spoken. Alternatively it drops a spoken line that sat under a drop. Spot it: a cue during a bar of music with no mouth moving, or a hole where the CTA was. Heavy beds confuse lipsync and transcription the same way. Fix: caption from a vocal-up mix, or duck the bed for the caption pass, or type the missed CTA as an authored overlay. Do not "turn on more accuracy." The engine is doing what it was asked: transcribe the file, including the junk in the file.

Offset versus drift

6. Uniform offset versus progressive drift. Both look like "the captions are late." They are not the same bug, and the same fix does not apply.

• Offset: every cue is early or late by about the same amount. Usual cause: you trimmed an intro after captioning, or the generator included pre-roll the picture no longer has. Fix: recaption the trimmed file, or shift the whole track by a constant if you still have a sidecar. A burned file has no track to shift. Recaption.
• Drift: the first cue is acceptable and later cues get worse. Usual cause: frame-rate mismatch (a 25 fps caption clock against 23.976 picture is the textbook case) or a generate that did not run at the clock you assumed. An offset will make the hook right and the closer worse. Fix: recaption against the actual file; do not slide.

Spot which you have by reading the first thirty seconds and the last thirty. If both are late by the same beat, offset. If only the end is late, drift. Caption timing is this clock, not the font.

A third clock complaint is a single late name in an otherwise tight file. That is not offset and not drift. It is a local alignment miss on that token. Recaption after correcting the word; do not slide the whole track to chase one cue.

Dumped paragraphs, and how to walk a file

7. Cue packing: one paragraph, no punctuation, unreadable at phone size. The engine heard a fast talker and emitted a 90-character block that covers the face and cannot be finished before the next sentence. This is not a transcription miss. The words can be right. It is segmentation plus reading speed plus line length, dumped into one cue.

Spot it on mute at phone size: you cannot finish the line. Spot it on the file: a cue that wraps to three lines, or that sits on screen for a blink. Fix: split at clauses, hold the cue long enough to read, cut the talker if the ad is faster than reading. Shrinking the type to keep a long sentence on one row makes it small and still too long. A "punchy" preset will not split a paragraph for you.

Walk every auto-captioned file in this order. Stop when you find the miss; do not restyle first.

1. Mute, phone, first two seconds. Placement and first-cue delay. Chrome covering the line is a safe area miss, not ASR.
2. Scan proper nouns, digits, slogans. Transcription.
3. Watch a two-shot or a noisy section. Crosstalk and music-as-words.
4. Check start and end. Offset vs drift.
5. Try to finish each line. Packing.
6. Then, and only then, argue about the preset.

Burned-in captions freeze whatever you missed. A sidecar would let you patch a word. A feed will not play the sidecar. Read, then burn. The AI caption generator is the transcribe-and-burn path; it will not know your stoplist unless a human applies it.

Language is not a separate seventh mode here because it is a setup error, not a model surprise: the captioner transcribes the spoken language. The wrong language code produces a bad transcript in the wrong language. Set the code to what is on the tape. A dubbed cut is a new caption job. That is configuration. The seven above are what still go wrong when the configuration is right.

Auto captions are a draft of the speech clock. The field guide is how you stop treating the draft as the master.

Read the next transcript before you burn it in Versely.

FAQ

Are these seven the same as "captions failed on mute"?

No. Mute is why you burn captions at all. These seven are why a burned file can still be wrong: the string, the speaker, the bed, the clock, or the packing. You can have a perfectly audible VO and still ship all seven.

Can I fix a burned miss by nudging in an editor?

Not if the captions are pixels. A wrong word, a packed cue, or a drift is a new render from the source. Sidecar workflows can patch a cue; short-form usually is not a sidecar workflow.

Which miss should I check first on an ad?

Names, numbers, and the first two seconds. A homophone in the hook and a line under the like button lose the auction before crosstalk in the testimonial ever matters.