---
title: Teach your mind how your project's words are spelled
date: 2026-11-18
summary: One small dictionary supplies speech hints, caption corrections and voice instructions without repeating the list in every app.
detail: One word list for hints and caption repairs
through: 499865e6c07a5cc1c6a715609915ce4b57177046
---

With this, your mind can keep its own vocabulary for speech, so technical terms come out consistently in captions.

Speech models do not know your project’s spelling. Repeating a correction in every voice prompt is tedious, and it does not fix text already transcribed. We now keep the terms and their known mistakes together in `words.md`.

Here is an invented project dictionary:

```markdown
- Mossdeck: moss deck
- Mossdeck Studio: moss deck studio
```

The left side is the spelling to keep. After the colon come observed wrong forms, separated by commas. Keep the list short. Never add a common word that might be correct in another sentence.

`words-parse.ts` turns the list into three things: cleaned, deduplicated transcription keywords; a transcript-repair function; and a paragraph for spoken-model instructions. `words.ts` imports the Markdown and exposes those shared helpers.

In `apps/voice-meet/app.ts`, the transcription session receives the canonical terms as keywords. `apps/voice-meet/backend.ts` supplies the repair function to the conversation handler, which corrects caption text server-side. Matching ignores case, respects whole-word boundaries and tries the longest phrase first. In our invented example, “moss deck studio” becomes “Mossdeck Studio”; a larger unrelated word stays alone.

The spoken AI gets the vocabulary differently: `apps/voice-meet/shared.ts` appends the paragraph to its instructions. `voice.ts` does the same for eligible voice sessions. That is a hint, not a guarantee of better hearing. Keep private vocabulary out of sessions that should not know it.

This layer needs no new package or service and makes no extra model calls. Our existing OpenAI live transcription and voice setup still needs an API account and a vault-held key; its audio charges still apply. Start a new speech session after publishing a dictionary edit; do not expect a running call to pick up new instructions.

The useful surprise is that hints and repairs solve different problems. Test both: try real audio for recognition, and unit-test casing, overlapping phrases and words that must remain unchanged. Our parser probe covered those text cases; it was not an audio-accuracy test.
