Why transcripts get names wrong, and how to stop it

Modern speech recognition handles ordinary sentences well. It has heard "we should push the deadline to Thursday" in countless variations and writes it down cleanly. What it has not heard is your account manager's surname, the working title of a product that does not exist publicly yet, or a compound term used by fifty people in your field.

When a recogniser meets a word it does not hold, it does the only thing it can: it writes something that sounds like it. The result is a transcript that is technically almost entirely right and still unusable, because the six words a reader actually scans for — who was there, which client, which product, which drug — are the six that came out garbled.

This is also why accuracy percentages are a poor guide. A transcript can be correct on almost every word and still fail, because errors are not distributed evenly across a page. They cluster on exactly the words that carry the meaning. Fixing them by hand is slow, and it is the kind of slow that makes people stop using transcription altogether.

A spelling glossary attacks that specific failure rather than accuracy in general. You are not asking the system to get better at English. You are telling it the handful of words it could not have known.

Special Names is a personal spelling reference attached to your account. You write down the words that matter to you, spelled the way you want them, and they are used as a reference for how to write those sounds when they occur in your recordings.

It is not find-and-replace. It is not a rewriting tool: it does not add words to a sentence, does not remove hedges or repetitions, and does not put a phrase in a speaker's mouth because the phrase is in your list. If a term was never said, it will not appear. It is not a rewriting tool either: it does not add words to a sentence, does not remove hedges or repetitions, and does not put a phrase in a speaker's mouth because the phrase is in your list. If a term was never said, it will not appear.

The boundary matters most in work where the transcript is a record of what someone actually said. In a clinical note, a client call or a research interview, a tool that quietly improves what was said is worse than one that spells poorly. Special Names is constrained on purpose: spelling, and nothing beyond spelling.

If you also want the wording cleaned up — filler removed, sentences repaired, headings and lists added — that is a separate choice you make per recording by selecting Smart Summary. The two are independent. You can have a strictly verbatim transcript with correct spelling, which is usually what regulated and evidentiary work needs.

Clinicians deal with drug names that differ by one syllable, condition names with several accepted forms, and abbreviations that are unambiguous in the room and ambiguous on paper. Adding the medications and conditions you actually prescribe and see, plus the names of colleagues you refer to, means those words come back spelled the way you entered them instead of approximated by sound.

Lawyers have a harder version of the same problem: case names, party names, counsel names and statute references all have to be exactly right, and they repeat across a whole matter. Adding them once at the start of a case means later recordings on that matter can draw on the same spellings, which matters more than it sounds — inconsistent spellings across a set of notes are hard to search later.

Consultants and salespeople live in client names and product names, often both from a company whose naming conventions no general recogniser has seen. Researchers face technical terminology and interviewee names together, and an interview transcript with mangled participant names is awkward to work with and awkward to quote.

And if you work in a language whose personal and place names are unfamiliar to a general recogniser, the effect is broader than a few terms. sAInote's interface ships in 11 languages — Arabic, Bengali, English, French, Hindi, Portuguese, Russian, Simplified Chinese, Spanish, Turkish and Urdu, with Arabic and Urdu rendering right-to-left — and speech recognition covers considerably more languages than the interface does. A glossary is how you close the gap on the names that remain.

Special Names is a Pro feature. Pro includes 360 minutes of audio a month, renewing monthly, along with all three processing modes and background handling of long recordings. If you are on Guest or Free, you get 15 minutes of audio once — not monthly, not recurring — with Quick Note and Smart Summary, and no glossary.

Recordings up to 20 minutes are processed immediately on every plan. Recordings from 20 minutes to 2 hours are processed in the background and are Pro only, with speaker separation and hierarchical summarisation for the long ones. Two hours is the hard cap per recording; anything longer is rejected. Uploads are capped at 64 MB.

Correct spelling pays off most when you come back later. Signed-in users get searchable note history across devices, and search matches both the raw transcript and the processed output. A name spelled three different ways across six meetings is a name you cannot find. A name spelled the way you wrote it in your glossary is one search away.

The glossary applies wherever you record: the web app in your browser, the iPhone and iPad apps, and the macOS app with its global hotkey Quick Capture on Control+Option+Command+N. The REST API is a separate path with two modes, raw and smart, authenticated with a secret key and metered per key; it stores neither audio nor text.