Vocabulary

Speech recognition is good at English and indifferent to your jargon. It will write hand of for handoff and lead code for LeetCode, and it will keep doing it. The vocabulary list in Settings → Vocabulary is how you fix that.

An entry has two parts, and the second is optional.

Preferred spellings

A term on its own is a preferred spelling, and it reaches both engines. Apple Speech receives your terms as contextual strings; the Whisper fallback receives them in the initial prompt it is given before transcribing. Either way it biases decoding toward them. It is a nudge, not a rule — the word becomes more likely, not guaranteed.

Good candidates are project names, library names, colleagues’ names, and domain vocabulary: VoxFlow, Tauri, Kubernetes, handoff.

Replacing a specific mishear

When a nudge is not enough, add a replacement: the text the engine actually produces, and what it should have been. This is applied after transcription rather than during it, so it is exact, and it applies whichever engine produced the text.

lead code      → LeetCode
cooper netties → Kubernetes

Replacements are applied to the raw transcript, and again after the rewrite pass, so a cleanup model cannot quietly undo them. Your vocabulary is also passed to the rewrite model so it knows the words are deliberate.

Learning from your corrections

In Settings → History, correcting an entry copies the fixed text and can offer to add the difference to your vocabulary.

VoxFlow can also notice corrections you make by hand. Learn from manual edits in Settings → Privacy is opt-in and off by default. With it on, VoxFlow inspects only the text range it just inserted, only for a short window afterwards, to spot a word you changed and offer to remember it. Nothing else on screen is read.

Suggestions you decline are remembered as dismissed, so you are not asked twice. Dismissed suggestions can be re-enabled at the bottom of the Vocabulary page.