Words and terms
A plain list of the names, products and jargon your team uses, plus terms learned from your Jira copy and your agent chats, and corrections for dictation.
Tell the transcriber what to expect before the meeting: product names, colleagues, acronyms, the internal word for a thing. It cuts the most annoying class of transcription error, and it takes one plain-text list. PrivateGoat also learns terms from your Jira copy and from what you type to your coding agents, and keeps spellings you remember; all of them can be reviewed and removed.
How to use it
- Open Settings → Dictation → Vocabulary → Words and Terms….
- Add your terms to your own list — one term per line. The window can open the text file for editing.
- Review the learned terms and remove any that are wrong. Learn terms from Jira switches them on; Refresh reads the Jira copy again. A removed term is blocklisted, so a later recompute does not bring it back.
- A spelling you remember after a transcript correction joins the same learned list. See Playback and corrections.
- Learned from your agent chats lists names and jargon from what you typed to Claude Code and Codex in the last 60 days — never the agents' replies — plus the names of your project folders. Learn terms from agent chats switches them on (on by default), Refresh reads again, and a removed word stays out. It needs Settings → Coding agents → Read coding-agent histories on this Mac; until then the section says how to allow it.
- Corrections (also Settings → Dictation → Vocabulary → Corrections…) holds "heard → meant" pairs for dictation: whole words, any case. Suggested from your recent dictations offers pairs with Accept and Dismiss; none applies until you accept it. Meetings are not changed by corrections.
Without Jira there is no Learned from Jira section; the window is about your own words and your chats.
The lists bias the recogniser — the same list is passed to Apple Speech's contextual strings and to Whisper's decoder context — so dictation and meeting transcription both benefit.
Settings
| Setting | Where |
|---|---|
| Words and terms | Settings → Dictation → Vocabulary → Words and Terms… |
| Vocabulary section | Settings → Dictation → Vocabulary |
| Corrections | Settings → Dictation → Vocabulary → Corrections… |
| Reading agent chats | Settings → Coding agents → Chats |
There is no per-meeting setting: the list applies to every transcription on the Mac.
What is stored
The files sit beside the workspace, so a redirected workspace root takes them along.
| File | Contents |
|---|---|
~/Library/Application Support/Sonavia/Workspace/Vocabulary.txt | Your own terms, one per line |
~/Library/Application Support/Sonavia/Workspace/LearnedVocabulary.json | The learned terms (from Jira, and spellings you remembered), the blocklist, the source field and a timestamp. The source is always written as jira, even for a remembered spelling. |
~/Library/Application Support/Sonavia/Workspace/LearnedFromAgentChats.json | The terms learned from your agent chats and their blocklist |
~/Library/Application Support/Sonavia/Workspace/Corrections.txt | Your dictation corrections, one "heard → meant" pair per line |
What leaves your Mac
Nothing. The lists are only passed to the local recogniser, and the agent chats are read on the Mac.
When it is off or fails
- A missing or unreadable file: the app treats it as "nothing learned yet"; transcription proceeds exactly as before.
- A removed learned term: it is added to the blocklist, so the next Jira-derived recompute does not resurrect it.
- Deleting
LearnedVocabulary.json: this forgets what was derived, not your own list. The two lists are separate. - An empty list: no effect at all — the model simply has no extra context.
- An unreadable
Corrections.txt: dictations are pasted unchanged and the window says so; Edit… opens the file.
Related
- Transcripts — where the list is applied.
- Dictation — the other engine that reads it.
- The local copy — where learned terms come from.