Tuning how fixes teach

Two settings decide how ozen learns from your fixes: how many learned words are hinted to the transcriber, and how many times a correction must repeat before ozen applies it by itself. ozen eval scores combinations of both so you can pick the best.

ozen eval --vocab 0,10,30 --repeat 0,1,2

It learns from half of a fixed set of spoken lines (as if you had fixed them) and scores each setting on the other half, best first:

• word error rate,
• English terms spelled right,
• word error rate on plain Hebrew,
• words invented on quiet noise.

By default the lines are synthetic (eval/cases.jsonl, spoken by macOS’s Hebrew voice). --real uses your own fixes instead. --fresh ignores cached results.

Runs are deterministic and cached, so a rerun prints the same table and a sweep only transcribes new settings.

The synthetic set is small and has one voice, so treat small gaps as noise and confirm a winner with --real once you have a few dozen fixes. To adopt a winner, set it as LEARN in src/fixes.rs and rebuild.

Comparing models on your speech

From ~/ozen, uv run eval.py [N] transcribes your last N real chunks with the stock, Hebrew, and Hebrew plus vocabulary setups side by side.