Hebrish for desktop v0.5.2

0.5.2: clearer Models tab; Sabich recommended.

• The model you dictate with is now outlined in green with an "In use" badge, so it no longer looks greyed out next to the others.
• The other downloaded models have a "Switch to …" button instead of "Use".
• Sabich, the complete model (the Hebrew words plus tech vocabulary, brand names and numbers), is listed first and recommended. On the first round of tests it was level with Jachnun on Hebrew words and better on tech dictation. Jachnun is still there; if one doesn't give you good results, try the other.
• Everything in 0.5.1: one copy of each model, updated automatically. See the 0.5.1 release.
• Ubuntu 26.04 .deb and a portable Windows zip. Not yet tested on Windows hardware. Downloads: hebrish.app/download.

Hebrish for desktop v0.5.1

0.5.1: one copy of each model, updated automatically.

• When a new version of a model comes out, Hebrish downloads it and deletes the old one once the new one has checked out, so each model is on your disk once. Old versions kept by earlier releases are cleared on first start (the one in use stays).
• Model updates install automatically by default (Settings → Model updates turns that off). The versions list on the Models tab is gone.
• Everything in 0.5.0: dictate into the window, spoken prompts, a level meter with a clip lamp, voice detection, Esc to discard and double-tap to start again, model cards with their own Download buttons, a choice of model folder, and a Diagnostics tab. See the 0.5.0 release.
• Ubuntu 26.04 .deb and a portable Windows zip. Not yet tested on Windows hardware. Downloads: hebrish.app/download.

Hebrish for desktop v0.5.0

0.5.0: dictate into the window, spoken prompts, a level meter, model cards, and text as you talk.

This release also carries 0.4.0, which was never published on its own.

• Text appears as you talk. Insert as you go is on by default: long takes are decoded at pauses while you speak, every ~10 seconds, so only the last few seconds are left when you stop. Long takes are split at real pauses, which cut the word error rate on ~50-second test takes from 16.0% to 1.0%.
• The Dictate tab records into a text box. Press Talk (or your hotkey) and the text lands in an editable box, as well as at your cursor. Output can be the cursor, the clipboard, or the text box only.
• Microphone pane: the mic in use (including which device "System default" is), a switcher, a level meter showing too quiet, good and loud zones and a clip lamp, and Test mic.
• Voice detection (on by default): a take with no speech isn't transcribed, so no stray "Yeah.", and long silences are shortened.
• Sounds: softer beeps, or short spoken prompts ("Recording", "Paused", "Transcribing", "Discarded", "Starting again"), or off. A volume setting and quiet hours.
• Keys: Esc discards a take; double-tap the toggle key to start again. One key per action for a macropad or pedals (Start, Stop, Pause, Retake, Discard).
• Models tab: a card per model with its own Download, Use and Delete, so you can download Jachnun, Sabich or stock Parakeet without switching. Choose the folder the models are stored in.
• File tab: drop an audio file to transcribe it.
• Diagnostics tab: inference speed on your computer (tokens per second, × real time, wait after you stop), with a benchmark button.
• History is now optional and off by default. Your existing settings get the new defaults once.
• Ubuntu 26.04 .deb and a portable Windows zip. Not yet tested on Windows hardware. Downloads: hebrish.app/download.

Hebrish for desktop v0.3.0

0.3.0: two hotkeys, a floating overlay, and a more finished look.

• Two hotkeys. A toggle key (tap to start, tap to stop; holding it is push-to-talk too, unless you turn that off) and a separate push-to-talk key. Either can be any key or a foot pedal, and either can be left unset. Your 0.2.x hotkey and mode carry over.
• Floating overlay (optional, like the Android floating mic): a small bar over every app with the timer, a level meter, Pause/Resume, Retake and Discard, then "Transcribing…" and "Inserted". Show it while dictating or always; drag it anywhere. It never takes the keyboard focus, so the text still goes where your cursor is. Turn it on in Settings or from the tray.
• Start-up screen while the model loads, and a redesigned About tab.
• Wayland: the .deb now pulls in the Qt Wayland plugin, so the window is native on GNOME and KDE. On GNOME the clipboard is set through XWayland (xclip), because GNOME doesn't let a background app set the Wayland clipboard. Raising the window from a second launch works on Wayland.
• Ubuntu 26.04 .deb and a portable Windows zip. Downloads: hebrish.app/download.

Hebrish for Android v0.6.0

0.6.0: a setup checklist, a side menu, simpler models, and a tidier voice keyboard.

• Setup checklist (home, and the side menu): every step in order, each ticking itself off: microphone, a model, the voice keyboard or floating mic, battery unrestricted, "Pause app activity if unused" off, and optional all files access. The same steps with pictures: hebrish.app/setup.
• Side menu (☰ on home) with every screen, Explore the lexicon, Contribute and a new About page. The app and model versions are at the bottom of home and the menu.
• Models as cards, each with its mark: Jachnun (English + Hebrew vocabulary) first and recommended, Sabich (Hebrew + tech vocabulary), stock Parakeet last. A new install uses Jachnun; an existing install keeps the model it has.
• Model location always shows where the models are, the default included. Back to the default copies them back into app storage.
• Voice keyboard: the mic sits at the top, with the Hebrish mark and a settings gear above it; comma and full stop keys beside space, delete and enter.
• Retake you can see: the mic and bar flash amber, say "↺ New take" and the mic spins back once. The mic pulses while recording.
• Optional beeps when a take starts (rising), stops (falling) and restarts (double pip); vibration on by default. Settings → Sound, microphone and buttons.
• The microphone in use is shown next to the timer. Choose one, or leave it on Automatic: a Bluetooth headset is always preferred.
• Start, stop, restart, pause and discard from outside the app: broadcast intents for Tasker, MacroDroid and Key Mapper, or learn a key from a Bluetooth shutter remote or keyboard. After upgrading, turn the floating mic off and on once so it can see buttons. Reference: docs/external-control.md.
• Floating mic: placeholder text such as WhatsApp's "Message" is replaced instead of left in front of your words (setting, on by default).
• Not yet tested on a device. arm64-v8a only. Android 9+. Downloads: hebrish.app/download.

Hebrish for Android v0.5.0

0.5.0: move the floating mic while you talk, and choose how it looks.

• Floating mic: drag it by any part, including the bar with the timer and buttons while you talk. Buttons still work as taps.
• Floating mic: press and hold, feel the buzz, then drag: holding still before moving no longer stops the drag during a take.
• Floating mic: new settings for the bar's background opacity while you talk, and for the mic's colour (six choices; it still turns red while recording).
• Not yet tested on a device. arm64-v8a only. Android 9+. Downloads: hebrish.app/download.

Hebrish for Android v0.4.0

0.4.0: faster, longer takes, a floating mic that works in more apps.

• Takes are transcribed while you talk, so the wait after Stop is short, and a take can run 20 minutes (was 90 s).
• Floating mic: finds the text field in apps built with Jetpack Compose, such as ChatGPT, where 0.3.0 often said there was no field or stayed hidden.
• Floating mic: a new "Keep it running" step in the guide lets Hebrish run unrestricted on battery, so phones like OnePlus don't stop it in the background.
• Floating mic: if no field had focus, tap one and then Insert. Finer placement: any height, anywhere across the screen, edge gap, size and idle opacity.
• The app has a screen per area: Models (download any or all, switch at once), Voice keyboard, Floating mic, Dictation, Install and updates (the F-Droid repository) and Help.
• Not yet tested on a device. arm64-v8a only. Android 9+. Downloads: hebrish.app/download.

Hebrish for Android v0.3.0

0.3.0: floating mic.

• A mic button that floats over every app (optional, an accessibility service). Tap it in any text field, talk, tap Stop: the text goes in at the cursor and your keyboard stays as it is.
• While you talk: Pause / Resume, Retake, Discard and a timer, next to the mic.
• Choose where it sits: drag it (it snaps to the nearer edge), or set the side, height and size on the setup screen, which moves it live. Option to show it only while a text field is focused.
• Long-press the idle mic to insert the last transcript again.
• Turn it on in the app's new step 4. On Android 13+ you may first need App info → ⋮ → Allow restricted settings.
• Not yet tested on a device. arm64-v8a only. Android 9+. Downloads: hebrish.app/download.

Hebrish for desktop v0.2.2

0.2.2: the model downloads from Hebrish's own CDN.

• The code-switching model (v1.0) now downloads from cdn.hebrish.app instead of Hugging Face, so it keeps working while the model repositories on Hugging Face need an access request. An existing v1.0 install is recognised and not downloaded again.
• Downloads resume and every file is checked against its SHA-256, as before.
• Terms of use: free for personal and non-commercial use; commercial use, including telephony, needs a licence. hebrish.app/terms.
• Ubuntu 26.04 .deb and a portable Windows zip (Windows still untested on real hardware). Downloads: hebrish.app/download.

Hebrish for desktop v0.2.1

0.2.1: model download fixed, terms of use, v1.0 results.

• Fixes the code-switching model download. The model's main repository on Hugging Face now needs an access request, so downloads and update checks failed. The app now downloads v1.0 from the open app build (Hebrish/…-sherpa-onnx-int8). Installed 0.2.0 apps also pick this up on their next update check.
• Terms of use in the About tab, and a note before the model download: the Hebrish models are free for personal and non-commercial use; commercial use, including telephony, needs a licence. Full terms: hebrish.app/terms.
• The About tab shows v1.0's benchmark: 5.1% word error rate on English with Hebrew words (stock Parakeet 11.3%), 73.3% Hebrew-term recall on the public benchmark.
• Ubuntu 26.04 .deb and a portable Windows zip (Windows still untested on real hardware).