Custom vocabulary

Speech models guess at unfamiliar words. Proper names, product names, team acronyms, and technical jargon are the usual casualties—"Hyprcore" becomes "hyper core", a teammate's name turns into something phonetic. Custom vocabulary fixes this by biasing the recognizer toward the terms you supply, before it ever produces a transcript.

This is different from custom words (covered at the bottom)—vocabulary steers recognition; custom words clean up afterward.

Add terms

Open Settings → Models → Speech and find the Vocabulary editor. Type a term, press Enter, and it appears as a chip. Remove one by clicking its ×.

Terms are stored in your vault at .vocabulary.yml, so they sync with the rest of your vault across devices.

You don't have to add everything by hand: meeting attendee names are biased automatically, and Hyprcore always biases the product-critical terms Hyper and Hyprcore—the first so Action Mode hears its wake word cleanly, the second so the app's own name transcribes correctly.

What it affects

Your vocabulary is fed to whichever engine is transcribing:

• Local models receive the terms as a compact glossary hint prepended to the transcription, nudging the model toward those spellings.

• Cloud STT (Deepgram) receives them as keyterm or keyword boosts, so the provider weights them higher.

A few practical limits:

• Up to 50 terms are used per transcription. Under that cap, your own terms and attendee names take priority over the built-ins.

• The glossary hint is length-bounded, so extremely long lists are trimmed to fit.

• Duplicates are ignored (case-insensitively), and your capitalization wins—add HYPRCORE and that's the spelling that gets biased.

Note: Biasing raises the odds a term is recognized; it isn't a hard rule. For a name the model still refuses to get right, pair vocabulary with custom words as a safety net.

Custom vocabulary vs. custom words

Hyprcore has two related but distinct features. Use whichever fits:

Feature	Where	When it acts	What it does
Vocabulary	Settings → Models → Speech	During recognition	Biases the model toward names and terms so it transcribes them correctly in the first place.
Custom words	Settings → Dictation → Behaviour	After the transcript	Corrects words that are often misheard or misspelled, matching similar-sounding output to your list.

Reach for vocabulary to teach the recognizer new names and jargon. Reach for custom words when a specific word keeps coming out wrong and you want it corrected after the fact.

What's next

• Speech models — Which engine is transcribing, and how bias reaches it.

• Action Mode — Why "Hyper" is always biased.