Vocabulary

Add names and phrases to help supported models recognize your terminology. Vocabulary is a recognition hint: it neither trains the model nor guarantees spelling. Voice and audio files share one list, with separate opt-ins.

Add and enable terms

1. Open Settings → Vocabulary, or Vocabulary in Voice’s or Audio file’s right options sidebar. Both edit the same list.
2. Enter one name or phrase per line. Spaces inside a phrase are preserved; blank lines and exact duplicates are ignored when sending hints.
3. Enable Voice transcription, Audio-file transcription, or both. Check each workflow’s support status; its help button explains model restrictions.
4. Choose Save. Changes apply to the next recording or file job, not active work.

The list and opt-ins stay saved when changing connections/models. Voice uses the same terms in completed and realtime mode where supported. Terms are stored locally and sent only to an enabled, supported transcription destination; review sensitive names before switching servers.

Upgrading from separate vocabulary lists

Freehand combines saved Voice, realtime, and file terms into one shared list. Previously enabled workflows may therefore send the combined list. Review the phrases and both workflow switches before the next transcription.

Check support and limits

Selection	How hints are sent	Request limit
Speaches with a compatible model	Recognition hotwords	2,048 UTF-8 bytes
Model/backend supporting context hints	Appended to existing transcription context	8,192 bytes, including context
NeMo-Speech.cpp v0.1.0 with explicit Nemotron 3.5 streaming profile	Vocabulary boosting in completed, file, and realtime requests	32 phrases; 128 bytes per phrase; 2,048 bytes total
Unsupported combination	Omitted; preference remains saved	No vocabulary sent

The saved list itself holds 16,384 UTF-8 bytes. Model request limits can be smaller; some characters use multiple bytes.

Over-limit lists stop the request

If the selected model cannot accept the complete list, Freehand stops before sending audio. Shorten the list or disable vocabulary for that workflow. Terms are never silently truncated.

For Nemotron, open Vocabulary tuning → Vocabulary strength (0–5). Start at 3 and review recognition. Stronger hints can produce incorrect matches; the server can cap or disable boosting. Voice and files share this strength.

Generic compatibility does not prove every model uses context hints. Vocabulary is separate from cleanup instructions, S1-mini context categories, and speech pronunciation; terms are not injected into those controls.

Review your list

Open lines to review below the editor for duplicates and lines exceeding the selected models’ phrase count, phrase size, or total budget. Line … highlights the phrase for editing. Context-based feedback includes existing context in the budget.

Feedback	Meaning
Exact duplicate	Sent once after trimming surrounding whitespace; the saved text is not rewritten
Case difference	A distinct phrase
Workflow off	Feedback only; it does not enable the workflow
Support check failed	Choose Try again; your draft stays in place

The list must fit the 16,384-byte storage limit before detailed line feedback is available. Save after correcting it.

Choose other recognition controlsUse subject context, temperature, or a model-specific option alongside vocabulary.