Dictionary Auto Learn
Turn corrections made after dictation into reusable vocabulary and word replacements.
What Auto Learn Does
Dictionary Auto Learn watches for edits you make shortly after VoiceInk inserts a transcription. It compares the inserted text with your corrected text and extracts small correction candidates.
VoiceInk asks your selected AI provider and model whether the edit contains a reusable correction of the same intended name or specialized term. The request includes the changed text and up to two nearby terms on each side so the model can recognize complete names and terminology.
- If accepted, VoiceInk adds the source and destination as a Word Replacement and adds the destination as one Vocabulary entry.
- If rejected, VoiceInk stores nothing.
For an accepted correction, the model selects the exact source and destination term from that small review window. VoiceInk validates that both selections were copied from the supplied text, then applies the storage behavior consistently.
Set Up Auto Learn
- Connect an enhancement provider in AI Models. You can use a cloud provider, a custom provider, Ollama, or another configured local option.
- Open Dictionary and click the gear button.
- Enable Auto Learn Dictionary.
- Select the provider and model used to review correction candidates.
This provider and model selection applies only to Dictionary Auto Learn. Changing it does not change the provider or model selected by any Mode for AI Enhancement.
How an Edit Is Processed
After VoiceInk inserts text, Auto Learn observes the same text field until you move focus away or 60 seconds pass. If you correct part of the transcription, VoiceInk:
- Compares the original insertion with the edited result.
- Extracts the changed source and destination with up to two nearby terms on each side.
- Discards candidates that are too large or ambiguous.
- Sends the remaining review windows to the selected AI provider for review.
- Adds every accepted pair to both Word Replacements and Vocabulary.
The review request contains only the bounded candidate windows. It does not include the recording audio, clipboard contents, selected text, screen captures, or the full contents of the field.
Nearby terms are used only to identify the complete reusable term. For example, when Wojciech was transcribed correctly but Szczęsny was not, the model can select Wojciech says me to Wojciech Szczęsny rather than creating an unsafe global replacement for says me.
What VoiceInk Should Learn
Good candidates are corrections that are likely to remain valid in future dictation:
| Transcription | Your correction | Expected result |
|---|---|---|
voice ink | VoiceInk | Accept |
post gray sequel | PostgreSQL | Accept |
api | API | Accept |
web hook | webhook | Accept |
Contextual changes should be rejected because they would be unsafe as permanent replacements:
| Transcription | Your edit | Why it should not be learned |
|---|---|---|
three files | four files | The quantity changed |
Tuesday | Wednesday | The date changed |
Python | Ruby | The intended subject changed |
deploy now | do not deploy now | The meaning changed |
An accepted voice ink to VoiceInk correction creates:
- A Word Replacement with
voice inkas the original andVoiceInkas the replacement. - One Vocabulary entry containing
VoiceInk.
Vocabulary entries can contain a word or a phrase. VoiceInk stores the complete destination as one entry rather than splitting it into separate words.
AI review can make mistakes. Review your Dictionary occasionally and remove any entry that should not be reused.
Ollama and Local Processing
To review candidates locally, connect Ollama under AI Models, make sure the Ollama service is running, and select an installed model in Dictionary settings. Candidate text stays on your Mac when the selected provider runs locally.
Cloud providers process candidate text according to their own data and privacy policies.
When Auto Learn Does Nothing
Auto Learn safely skips an edit when:
- The edited field cannot be read through macOS Accessibility APIs.
- The edit is too large or its boundaries are uncertain.
- The model rejects the candidate.
If AI Review Fails
If no compatible provider is configured, the selected provider or model is unavailable, or the model returns an invalid response, VoiceInk keeps pending correction candidates and shows a warning in Dictionary. Open Dictionary settings to see the error, choose another provider or model if needed, and click Retry.
A later correction also triggers another review attempt. The warning clears after the next successful AI review. VoiceInk sends all currently pending candidates together and keeps the 100 most recent unique pending candidates so repeated provider failures cannot grow the queue indefinitely.
Manage Learned Entries
Automatically learned entries appear in the regular Vocabulary and Word Replacements sections. You can edit or remove them in the same way as entries added manually.