Phona 1.5.0

Fifteen changes since 1.4.0. The speech model changed, the hotkey changed, and the correction stage lost three of its four modes.

Speech runs on Parakeet now

Parakeet TDT 0.6b v3 replaces Whisper large-v3-turbo. The two tie at 2.45% word error rate on the LibriSpeech test-clean sample and tie again at 12 of 14 on planted grammar errors, but Parakeet runs at RTF 0.030 against 0.107, so the speech stage costs about a third of what it did. It also returns an empty string on silence and noise where Whisper invented "thanks for watching".

This reverses the call made in 1.2.0. The original comparison used twelve sentences and was too small to carry the decision.

Parakeet takes neither a language nor an initial prompt, so the dictionary hint does not reach it. Run ./switch-model.sh whisper if you need that hint.

Tap Option, do not hold it

Tap the left Option key to start, tap again to stop. Holding no longer records, so a hand resting on the key cannot open a five minute recording. Every Option shortcut keeps working, and the right Option key is a plain modifier.

One correction mode

Four prompts behind one setting became one. Two of the four had no grammar rules at all, so moving the setting quietly changed how well your grammar was fixed. Every guard now runs on every correction.

Corrections stopped being thrown away

The guard treated a hyphenated compound as one word, so "AI-assisted" matched neither "AI" nor "assisted" and the correction was discarded. That was the single largest cause of rejection, 8 of 22 on the 8-bit model. Dictations that came back with no grammar correction at all halved.

The grammar model is now measured

Qwen3-4B at 8-bit was picked as a default and went unmeasured for a month. Four models were scored over 402 corrections drawn from real dictation. It won on precision: 95 repairs that touched nothing else against Gemma 4's 68, and 96 stray edits against Gemma's 231. Gemma repairs more errors and was not chosen, because for dictation an unwanted edit is worse than a missed one.

Score it yourself with tests/eval_correction.py.

Also

• Long dictations are split and rejoined rather than truncated
• Spoken layout commands: "new paragraph", "new line", "bullet point"
• A mail style that expands contractions, and a chat style that drops the closing full stop
• Literal replacements applied before correction, for names the model mishears
• Audio retention is configurable and off by default
• A weekly audit into audit-latest.md

Install

Download Phona-1.5.0.dmg, drag Phona to Applications, then run ./install.sh from this repo for the speech engine.

The app is signed ad-hoc rather than with a paid Apple Developer ID, so macOS warns the first time. Right click, Open, then Open again. Once only.

Apple Silicon and macOS 14 or later.