What changed, and when.

In testing · research branch

Two engines that change the language question.

• Supertonic 3 (31 languages) and OmniVoice (646 languages) now run behind the read-aloud daemon as first-class engines. Supertonic is real time on a bare CPU; OmniVoice is real time on an Intel Arc iGPU. Both rated very good by a native Dutch listener.
• A fallback ladder keeps reading working everywhere: OmniVoice, then Supertonic, then Piper.
• omalexia voice advise reads your machine (CPU, RAM, GPU, compute stack) and recommends the best engines it can sustain.
• The multilingual speech research written up in full: see the research page.

These are on a research branch and not yet in the beta installer; they land here when they do.

30 August 2026

Point at it and listen.

• Read from the word under the pointer: hover a paragraph, press Alt+F10, and reading starts there. No selection needed.
• Alt+F10 now works in every window: when an app offers no text through any channel, OCR completes the grab ladder as the last resort.
• The Ghostty terminal joins the exact grid, so read-along highlighting is character-precise there; Alacritty is covered by the fallbacks.
• The endgame written down: bet on the standard accessibility interfaces, so per-app adapters can retire as apps catch up.

28 August 2026

Ask the window for its text.

• The accessibility tree becomes a text backend: any app that exposes its text the standard way is asked for it directly, no pixels involved.
• The foot terminal gets exact highlighting the same way: character positions from the terminal itself instead of reading the screen.

26 and 27 August 2026

The words light up as they are spoken.

• Read-along highlighting born: while text is read aloud, the word being spoken is highlighted in the actual text on screen, in any app.
• Locating words on screen made fast enough for dense screens, then faster: selections are found by their highlight colour instead of full OCR.
• The marker follows when you scroll, stays in sync with the audio you actually hear, and no longer flickers or skips words.
• Mark the word, the sentence, or both; browser selections take the fast path.
• Reading survives a daemon restart mid-text.

25 August 2026

One calm workspace, eight Kokoro languages.

• Reading mode moves the window to one clean workspace and restores everything on exit.
• Kokoro switched on for all eight languages it actually speaks; only our own settings had limited it to English.
• Measured Chatterbox Multilingual end to end and found the real bottleneck (the audio decoder, not the language model); planned the path to commercial-grade voice quality.

24 August 2026

Day one: the whole first cut.

• The read-aloud daemon: two neural voices kept warm, speech about 0.1 s after the key press, sentence streaming, instant stop, language detected from the text. Registered as the system voice, so the browser's reader mode and the screen reader speak with it too.
• Read in any language: voices download on demand, a language selector, wider detection.
• Dictation tuned: the multilingual Parakeet model, audio ticks, filler-word cleanup in English and Dutch, a personal word list.
• The calm screen: reading mode on one key, a paper tint, the Atkinson Hyperlegible reading font desktop-wide (OpenDyslexic one command away), bigger text everywhere.
• The bar widget and its panel: read, stop, speed, voices, dictation, reading mode, font and text size in one place, plus plenty of panel polish.
• Reading speed up to 4x, the stall between sentences removed, speed changes apply mid-text.
• Measured the speech stack and the NPU; recorded the findings and a roadmap.
• This site went up; the project moved into its own repository under thefreshoffice, MIT licensed, contributors welcomed.