Mozhi v0.3.1

Mozhi 0.3.1

A fix release for v3. Everything new in 0.3.0 still applies: edit while you speak, emoji in the pill, errors that keep your words.

Fixed

• A phrase could go missing from a dictation. If the pill was still loading when Mozhi committed your first phrase (most likely right after launch, or on a busy machine), that phrase could be dropped from what was inserted. Phrases are now numbered: the pill notices a gap and resyncs, and Mozhi never accepts a draft that skipped one.
• "Mozhi quit unexpectedly" on quit. whisper.cpp's Metal backend aborted while shutting down because the speech model was still loaded. The model is now released before Mozhi exits.

For testers

mozhi --selftest speech.wav --selftest-app com.apple.TextEdit dictates a recording into the given app, and pastes only if that app really is in front (otherwise it copies), so an automated run never types into the wrong window.

Install

macOS (Apple Silicon, macOS 12+): download Mozhi-macOS-AppleSilicon.dmg, open it and drag Mozhi to Applications. Mozhi isn't notarised by Apple yet: the first time, right-click it in Applications → Open → Open. Then allow Microphone and Accessibility. Updating from an earlier version? Remove the old Mozhi entry in System Settings → Privacy & Security → Accessibility and add the new app; macOS treats each unsigned build as a new app.

Windows (10/11, 64-bit): download Mozhi-Windows-x64-setup.exe (or the .msi) and run it. SmartScreen may say the app is unrecognised: More info → Run anyway. The Windows build hasn't been tested on a real PC yet; please report anything odd.

The first run downloads a speech model (about 265 MB for the recommended Tamil model). After that everything runs offline.

Mozhi v0.3.0

Mozhi 0.3.0 · மொழி v3

Edit while you speak

Click into the text while Mozhi is still listening. Move between words, fix a name, add a line;
Mozhi keeps listening, and every phrase you say next goes in where your cursor is. The words it's
still hearing show on a live line underneath. ↵ inserts, ⇧↵ adds a line, esc cancels.

Emoji, without leaving the pill

• A 😊 button in the pill opens an emoji picker: an everyday row (🙏 😊 😂 ❤️ 👍 🔥 🎉 🪔 🍛 ☕),
  your recent picks, and every emoji by group.
• Search in English, Tamil or Tanglish: heart, இதயம், sirippu, sapadu, vanakkam.
• Say it: "heart emoji", "thumbs up emoji", "இதயம் எமோஜி", "சிரிப்பு எமோஜி". Turn it off in
  Writing → Voice emoji.

Errors that keep your words

• If a transcription fails, your recording is kept. Try again transcribes it once more and shows
  the result for you to check before it's inserted.
• Every error now offers the fix that matches it: microphone or Accessibility settings, Models,
  or Copy text.
• Report (in the pill and in About) opens a GitHub issue already filled in with your version,
  system, model, the error and recent log lines. Nothing you dictated is included, and nothing is
  sent until you submit it.
• Fixed: "failed to encode" errors when you finished speaking while a live-preview pass was still
  running on the GPU. Inference is now serialised and a failed final pass retries once.

Under the hood

• Phrases are committed at natural pauses and transcribed at full quality, so long dictations
  stay accurate and the final pass after ⌃⌃ only has the last few words to do.
• mozhi --selftest speech.wav runs a complete dictation from recorded audio, for release checks.

Install

macOS (Apple Silicon, macOS 12+): download Mozhi-macOS-AppleSilicon.dmg, open it and drag Mozhi to Applications. Mozhi isn't notarised by Apple yet: the first time, right-click it in Applications → Open → Open. Then allow Microphone and Accessibility. Updating from an earlier version? Remove the old Mozhi entry in System Settings → Privacy & Security → Accessibility and add the new app; macOS treats each unsigned build as a new app.

Windows (10/11, 64-bit): download Mozhi-Windows-x64-setup.exe (or the .msi) and run it. SmartScreen may say the app is unrecognised: More info → Run anyway. The Windows build hasn't been tested on a real PC yet; please report anything odd.

The first run downloads a speech model (about 265 MB for the recommended Tamil model). After that everything runs offline.

Mozhi v0.2.0-beta.1

Mozhi · மொழி v0.2.0 beta

Open-source, on-device Tamil voice typing. Double-tap Control, speak, and the text lands in the app you're in.

What's new

• Tamil model by default. New installs download Tamil Small, Whisper fine-tuned on Tamil by IIT Madras SPRING Lab: the most accurate Tamil in the catalogue, at about 265 MB instead of 575 MB. If you already have a model, Mozhi keeps using it until you pick another in Models. Dictate a lot of English? Choose Whisper Large v3 Turbo there.
• Edit before you insert. Click the text in the voice pill (while speaking or after) and fix it right there, with a cursor and normal editing. Enter puts it into the app you were typing in, Shift+Enter adds a new line, Esc cancels. Your app keeps its place the whole time.
• Prefer to always check first? Turn on Pause to edit before inserting under Shortcut, and every dictation stops in the pill for review.
• Three lines, always. The pill shows at most three lines; longer text scrolls, with older lines fading out at the top.
• Clipboard button in the pill. Saves what you said (to History and the clipboard, nothing is lost mid-sentence) and opens History, which lists everything you've dictated.
• Copy all in History copies every dictation at once, oldest first.

Install

This is an unsigned beta for testing.

macOS (Apple Silicon, macOS 12+): download the .dmg, open it and drag Mozhi to Applications. The first time, right-click Mozhi in Applications and choose Open, then Open again (it isn't notarised yet; if macOS still refuses, use Open Anyway in System Settings → Privacy & Security). Allow Microphone and Accessibility when asked.

Windows (10/11, 64-bit): download the -setup.exe (or the .msi) and run it. If SmartScreen warns, click More info → Run anyway.

Mozhi downloads a speech model the first time (about 265 MB for the recommended Tamil model). After that everything runs offline.

Found a problem? Report it here.

Mozhi v0.1.0-beta.1

Mozhi · மொழி beta

Open-source, on-device Tamil voice typing. Double-tap Control, speak Tamil, Tanglish or English, and the text lands in the app you're in.

This is an unsigned beta for testing.

macOS (Apple Silicon, macOS 12+)

1. Download the .dmg, open it and drag Mozhi to Applications.
2. The first time, right-click Mozhi in Applications and choose Open, then Open again. (It isn't notarised by Apple yet.) If macOS still refuses, open System Settings → Privacy & Security and click Open Anyway.
3. Allow Microphone and Accessibility when asked. Accessibility is what lets Mozhi hear the double-tap and paste for you.

Windows (10/11, 64-bit)

1. Download the -setup.exe (or the .msi) and run it.
2. SmartScreen may warn that the app is unrecognised: click More info → Run anyway. (It isn't code-signed yet.)

First run

Mozhi downloads a speech model (about 550 MB for the recommended one) the first time. After that everything runs offline.

Found a problem? Report it here.