Teach Lispr your words

Every dictation user has a private list of words their tool keeps getting wrong. Your boss's surname. The product you ship. The framework you live in. The list is short, the errors are the same every time, and re-typing them is the slow tax that makes voice feel not quite worth it. There is a way to teach the model these words, and the rest of this is about which words to pick and what teaching them actually does.

What the hint is, and is not

Vocabulary is a list of correct spellings you teach the app. Before every dictation, the app passes that list to the transcription model as a hint — a short note that says, roughly, "if any of these words sound plausible in what you are about to hear, prefer this spelling."

The hint biases the model. It does not override it. The model still listens to your audio and still picks the words it thinks are most likely. Vocabulary tilts the scale toward your spelling when the sounds are close. This is the same mechanism that explains why technical terms are hard in the first place — the model leans on what it has seen before, and rare proper nouns are exactly the words it has not seen. The Vocabulary gives it a few it has now.

A useful nudge. Not a rule.

How to open it and what you see

Click the Lispr icon in the menu bar at the top of your screen and choose Vocabulary… from the dropdown. There is no keyboard shortcut and no Settings tab.

You see two toggles and a table. The feature ships off. Flip Enable on and the Terms section appears. The second toggle, Auto-grow the vocabulary, lets the app scan your recent dictations and add brand-words it thinks you would want. You can leave it on and forget it, or turn it off and curate by hand.

The Terms table has two columns: the correct spelling on the left, and a small count of how many different ways the app has misheard it on the right. A person icon marks entries you pinned manually.

Adding and removing entries

Click Add. A small sheet appears with two fields: how the app currently hears the word, and how it should be spelled. So for me: carpoosian → Karpushin. Press Return. One pair at a time, no bulk paste. One at a time is the whole model.

To remove an entry, click the X on its row and confirm. The app remembers the deletion, so the auto-scan will not re-suggest the same word later. To fix a spelling, click the pencil and edit in place.

The interesting question is which words to put in.

The four kinds of words that pay off

Vocabulary helps most with words the model has no way to guess. Four categories cover almost everything worth adding.

Names of people you dictate often. Your boss, your closest colleagues, the three or four clients whose names come up weekly. Non-English names dictated in English sentences are the strongest fit — "Siobhan", "Karpushin", "Myroslav". The model has weak priors for these by definition.
Names of companies, products and projects. "Codebridge" (otherwise: code bridge), "OpenClaw" (otherwise: open claw), "Lispr", and the internal project code only you and a few teammates use. Made-up brand names are exactly what the model was never trained to spell.
Acronyms you pronounce as words. "SaaS" (otherwise: sass), "AWS", "GPT", "RPA". These are particularly worth adding if you dictate in a language other than English and the acronym keeps the English pronunciation — that combination scrambles transcriptions in predictable ways.
Foreign-language proper nouns inside otherwise-English sentences. "Monobank", "Sberbank", "Nova Poshta", a Japanese product name dropped into an English email. The model handles each language reasonably well; it handles the switch poorly. A hint helps it land on the right spelling instead of a phonetic English approximation.

A note on what not to add. If the only thing Lispr gets wrong is the capitalization — "monobank" instead of "Monobank", "google" instead of "Google" — Vocabulary will not help you. The app filters casing-only corrections out before sending the hint, on purpose. Casing is a job for your editor. Don't add common English words, don't add full sentences, and don't put the correct spelling in both fields.

Punctuation, apostrophes and accented characters are honoured — type the name exactly as you want it written, so "O'Brien" and "Müller" land correctly. Multi-word entries like "Nova Poshta" go in as one row. Cyrillic, Hebrew and CJK entries work too — add "Карпушин" and it will land in Russian dictation. One Vocabulary covers every language you dictate in; there is no per-language split.

How the app learns on its own

With Auto-grow the vocabulary on, the auto-scan looks at your recent dictations at most once every 24 hours, and only when you have dictated at least about ten new phrases since the last scan. New rows appear directly in the Terms table. There is no inbox of pending suggestions to approve; your review surface is the table itself. Read the new rows, delete the wrong ones, edit the ones that are close.

If you have just dictated a meeting full of new names and do not want to wait, hit Search now. It runs the same scan immediately.

Be honest about what auto-learning can do. It is good at noticing recurring brand-words. It cannot read your mind about how you spell your nephew's name. For anything that matters — a client name in a contract, a product name on a press release — add it by hand. Manual entries also carry more weight in the hint, which matters for the next section.

What it cannot fix

The hint biases the model; it does not override physics.

A word still has to sound roughly like what you said. If you mumble a name, Vocabulary will not save it.
A word that has never been in the Vocabulary and that the auto-scan has not yet caught will still be misspelled.
A name that competes with a common real word will sometimes lose. "Xero" in your Vocabulary will not always beat "zero" in a sentence about counts.
Casing-only fixes are out by design.

There is a budget per dictation

There is a hard ceiling on how much the app can whisper to the model before each dictation — a 500-byte cap on the hint string. In practice, that is roughly forty to sixty short English names like "Codebridge, Xero, OpenClaw" — fewer if you dictate in Cyrillic, Greek or CJK scripts, where each letter costs two or three bytes. You can store far more entries than that, but only the highest-confidence subset that fits the budget ships on any given call. The tail is silently dropped.

Two things follow. Entries you added by hand always rank above auto-discovered ones, so the surest way to guarantee a word is honoured is to add it yourself. And a lean list of carefully chosen entries protects you more reliably than a sprawling one — a focused forty-name Vocabulary covers most of the gain available; three hundred entries do not add three hundred entries of accuracy.

Curate. Keep the list short, deliberate, and made of words you actually say.

Where the list lives

The Vocabulary is stored on your Mac, as a plain file in Lispr's application support folder. There is no account, no iCloud sync, no cross-device sync. The list lives on the machine you built it on, and if you set up Lispr on a second Mac you will rebuild it there. That is deliberate.

Your correct spellings are sent with each dictation when the feature is on and you have at least one qualifying entry — that is the hint. If the auto-scan is enabled, recent dictation snippets are also sent for the brand-word analysis, each capped at 500 characters, along with your existing vocabulary keys so the server will not re-suggest words you already know. Deleting an entry tells the scan not to re-suggest it.

A starter list for week one

If you are looking at an empty Vocabulary and wondering where to begin, three rows on a napkin will do.

The five to ten people whose names you say or write each week — your manager, your closest teammates, the clients you email most.
The two or three product, company or project names you live inside — yours, your employer's, the codename for the thing you are shipping this quarter.
The handful of technical terms or acronyms you have already noticed the app garbling.

Add those by hand. Turn on Auto-grow the vocabulary and let it pick up the long tail over the next week. Come back in seven days and prune anything wrong.

Lispr keeps that list on your Mac and sends it alongside each dictation as a hint. Press your Lispr key, dictate the names you have taught it, and trust the first forty entries to do most of the work.