· Everpop
Speech recognition is tuned to get most words right, and a proper noun is not most words. Names are rare in training data, phonetically close to commoner words, and arrive with no context to disambiguate them. So the system quietly swaps in something that sounds about right. Overall accuracy stays high. The surname is wrong.
Why does a high-accuracy transcript still get the name wrong?
Because word error rate is an average, and it counts every word the same. A dropped "the" and a butchered surname cost identical points. Researchers working on pulling named entities out of speech said it flatly: ASR systems "are generally tuned in order to get the lowest word error rate on a validation corpus, but this metric is not optimal to the NER task", and it "does not distinguish between errors on verbs or proper nouns while such errors do not have the same impact" (arXiv 1805.12045).
The gap behind the average is not small. A 2025 paper on recognising named entities in classroom speech benchmarked a widely used open speech recognition system at "a WER of 2.7% on the standard Librispeech clean corpus" — and then measured "a WER of 28.9% for persons and 19.6% for organizations" on a benchmark built around real conversational context (arXiv 2506.10779).
On clean read speech, near flawless. On human names in context, roughly one in four goes wrong. Both true at once.
What is actually going wrong inside the transcript?
Three mechanisms, and they usually stack.
The name was never really in the vocabulary. The same 2025 paper notes that named entities "relate to events, persons' names, products, etc. which often depend on local context and change over time". A surname, a six-month-old product, a founder nobody outside the industry has heard of: none of these are stable, high-frequency words. They are the long tail, and the long tail is where recognition thins out.
It reaches for the commoner word that sounds the same. The same paper describes the failure directly: speech recognition "may incorrectly capture named entities due to similar-sounding words or uncommon entities in training data". Google's own speech documentation frames the general version of this as biasing toward "weather" over "whether" (Google Cloud, model adaptation). A rare name loses that contest every time, because a common word is, by construction, more likely.
There is nothing to disambiguate against. The system hears audio. It does not have your guest list, your show notes, or your sponsor's brand book. A human editor hearing an unfamiliar surname stops and checks. The transcript does not stop. It commits, confidently, in writing.
A 2026 write-up on transcribing NBA commentary logged "Kristaps Porzingis" coming out as "Christmas Por Zingas" and "Devin Booker" as "devon bucker", with overall word error rate sitting at 0.217 — and found that "Names replaced with phonetically similar common words" made up "35% of errors", the largest category (arXiv 2602.18966).
"Christmas Por Zingas" is very funny in a research paper. It is markedly less funny burned into the frame, in a large friendly font, beside the face of the man it belongs to.
Which words break most often?
| Word type | Why recognition struggles | Typical failure |
|---|---|---|
| Surnames outside the dominant language | Rare in training data; spelling not predictable from sound | Split into two ordinary words |
| Brand and product names | Deliberately coined, often a respelled ordinary word | Reverts to the ordinary spelling |
| First names with variant spellings | Every variant sounds identical | The most common spelling, not the one on the passport |
| Acronyms and initialisms | Letter runs resemble ordinary words | Rendered as a word, or half spelled out |
| Domain jargon | Low frequency, phonetically near a common term | Swapped for the common neighbour |
| Place names | Local pronunciation drifts from local spelling | Nearest familiar place, or a phonetic guess |
None of these are edge cases. They are the highest-stakes characters in the clip.
How do you stop captions spelling names wrong?
Tell the system the words before it listens
Several transcription services accept a list of terms to weight in advance. AssemblyAI documents keyterms prompting as accepting "up to 1,000 words or phrases (maximum 6 words per phrase)" to "improve transcription accuracy for those terms and related variations or contextually similar phrases" (AssemblyAI docs). Deepgram documents keyterm prompting for "up to 100 important terminology, product and company names, industry jargon, phrases and more", including "Proper nouns: Names, brands, titles with appropriate capitalization" (Deepgram docs). Google's model adaptation helps the system "recognize specific words or phrases more frequently than other options that might otherwise be suggested", aimed at words "that are rare in general language use (such as proper names or domain-specific words)".
Google's documentation caps the practical boost value at 20 and warns that pushing higher trades missed words for wrongly inserted ones. It is a thumb on the scale, not a guarantee: better odds on the words you named, nothing for the ones you forgot.
Review it before it is public, not after
YouTube's own guidance on automatic captions is blunt: they "might misrepresent the spoken content due to mispronunciations, accents, dialects, or background noise", and creators should "always review automatic captions and edit any parts that haven't been properly transcribed" (YouTube Help). That applies with more force to burned-in captions, which cannot be quietly corrected later — the pixels are the pixels. Word-timed captions and muted autoplay covers that trade-off.
This is the part Everpop is built around: nothing publishes until you approve it, so a mangled surname is a rejection rather than an apology. Re-renders do not touch the transcript — three free re-renders per clip cover another pass at the cut, not the caption text. No transcription step reliably spells an unfamiliar name right first time, and Everpop does not pretend to.
Correct it once, in a file you keep
Fixing the same name per clip, per platform, is how a small problem becomes a standing tax. Fix it once in a subtitle file instead. SubRip (.srt) is on YouTube's supported list, and the requirement is simply that "The file must be in plain UTF-8" (YouTube Help) — plain text, editable anywhere, portable to the next platform. Everpop's editor handoff exports FCPXML, EDL and SRT for exactly this reason: correct the text in your own tools and the corrected file stays yours. Exporting clips to Premiere and DaVinci walks through it.
What should you check before you publish?
Run this before anything with a name in it goes out.
- Write the name list first — guests, brands, products, places, acronyms — spelled as you want them on screen.
- If your transcription step takes a keyterms or custom-vocabulary list, paste that list in before it runs.
- Read the words on screen, not in a text pane, and check every frame a name appears in — including the ones it appears in once.
- Search the transcript or SRT for every name on your list. Something that sounds close but reads wrong is the failure mode, not a typo.
- Check the title and description too. If you pasted from the transcript, the error travelled.
- Correct it in the SRT and keep that file for the next platform.
- If the name is still wrong, do not publish. Send it back.
A mangled name is not a small quality issue. It is a message to the one person in the audience guaranteed to notice, saying nobody watched this before it went out. Easy to prevent, awkward to explain.
Frequently asked questions
- Why do captions spell names wrong when the rest of the transcript is fine?
- Because accuracy is measured as an average across all words, and names are a tiny fraction of them. Research on named entities in speech recorded a 2.7% word error rate on a clean read-speech corpus alongside 28.9% on person names in a context-dependent benchmark. The average looks healthy while the names quietly fail.
- What is custom vocabulary or word boost, and does it solve this?
- It is a list of terms you supply before transcription so the system weights them more heavily. AssemblyAI documents up to 1,000 words or phrases; Deepgram documents up to 100 terms including proper nouns and brand names; Google calls it model adaptation. It improves the odds on words you name in advance. It cannot help with words you forgot to list, and Google warns that boosting too aggressively can cause terms to be inserted where they were not spoken.
- Can I fix a burned-in caption after publishing?
- Not in the file you already published. Burned-in captions are part of the video frames themselves, so correcting one means re-rendering the clip and posting it again as a new upload. That is the whole argument for catching name errors at review rather than after.
- Does correcting an SRT once mean I never have to fix it again?
- For that clip, largely yes. SubRip (.srt) is plain UTF-8 text with timings, accepted by YouTube among other platforms, so a corrected file can be reused wherever you post that clip. New clips still need their own pass, which is why a running name list is worth keeping.
- Does Everpop let me edit the caption text to fix a name?
- Everpop's role here is the approval gate, not a transcript editor: nothing posts until you approve it, so a clip with a mangled name can be rejected before it is public. For text corrections, the editor handoff exports FCPXML, EDL and SRT, so you can fix the wording in your own editor or caption tool and keep the corrected file.
- Which names are most at risk?
- Anything rare, recently coined, or phonetically close to a common word: surnames outside the audio's dominant language, brand and product names built from respelled ordinary words, first names with several accepted spellings, acronyms, domain jargon, and place names whose pronunciation drifts from their spelling.