voculinux/docs/UPDATE.md

# Updating Voculinux

How to upgrade an existing install, plus release notes for the current series.

## Quick update

Void package installs (the signed xbps repository registered in [INSTALL.md](INSTALL.md)):

```bash
sudo xbps-install -Su voculinux
```

A package-managed install makes no release check of its own, and Settings → About says so.
Updates arrive through xbps only.

Source checkouts (developers) check this project's own release server, the Gitea repository at
`git.bongbetic.com/xavierk/voculinux`, in the background about every six hours and shortly
after startup. When a newer build is on your selected channel, the tray menu gains an
**Update Available...** entry and Settings → About shows a **New** badge; open that page for
release notes and download links. Updates are not installed automatically. Re-run the
installer as below.

### Source checkout

```bash
cd voculinux
git fetch origin
git checkout v0.18.5
./install.sh
```

Latest development tree:

```bash
cd voculinux
git pull origin main
./install.sh
```

On an existing install (a virtual environment plus `~/.config/voculinux/config.json`),
`./install.sh` reports the saved configuration and defaults to **Upgrade (keep current
settings)**: the configured engine stays selected, `config.json` is never rewritten, the
virtual environment is reused, downloaded models are kept, and pywhispercpp is rebuilt only
when this release pins a new version or GPU support is missing. Pick the interactive or
automatic modes instead if you want to reconfigure the install.

For a scripted, non-interactive upgrade:

```bash
./install.sh --upgrade
```

`--upgrade` on a machine with no existing install falls back to a normal automatic installation.

### Check your version

```bash
voculinux --version
# or
python3 -c "import vocalinux; print(vocalinux.version.__version__)"
```

### Update problems

Clean reinstall (keeps config and models by default). Void package:

```bash
sudo xbps-remove voculinux
sudo xbps-install -S voculinux
```

Source checkout:

```bash
./uninstall.sh --keep-config --keep-data
./install.sh
```

If an old process is stuck, stop via the tray or the PID in the instance lock file (do not use
`pkill -f voculinux`; it can kill unrelated processes):

```bash
kill "$(tr -d '[:space:]' < "${XDG_DATA_HOME:-$HOME/.local/share}/voculinux/instance.lock")"
# If you use IBus injection:
kill "$(tr -d '[:space:]' < "${XDG_DATA_HOME:-$HOME/.local/share}/voculinux-ibus/engine.pid")"
voculinux
```

Missing system packages: see [INSTALL.md](INSTALL.md) or [DISTRO_COMPATIBILITY.md](DISTRO_COMPATIBILITY.md).

---

## What's New in v0.18.5

0.18.5 is a **patch** on the stable line. Default engine is still whisper.cpp. It closes the reports filed after 0.18.4: a stop pressed while a recording is starting can no longer leave the microphone recording forever, microphones that stall mid-read can no longer wedge dictation, keyboards survive suspend and the GNOME lock screen, Ctrl+mouse-wheel zoom works again, and the installer learns an upgrade mode that keeps your settings.

### 0.18.5 highlights

| Fix | Description |
|---------|-------------|
| **A racing stop always wins** | A stop arriving while a recording is still starting is no longer overwritten by the start, so the microphone can never keep recording with no way to halt it (#1023 by @aagprojectsteam-max; fixes #1022 reported by @aagprojectsteam-max) |
| **Stalled microphones cannot hang Stop** | Capture reads happen outside the buffer lock, so a mic that stalls mid-read (a USB or Bluetooth hiccup, or the audio server reconfiguring the device) no longer leaves dictation stuck in PROCESSING (#1023; supersedes the approach in #974 by @neuroborus) |
| **Keyboards survive suspend and lock** | The evdev forwarder releases its EVIOCGRAB when forwarding cannot run, so keyboards no longer go dead at the GNOME lock screen after a suspend/resume (#1025; fixes #1021 reported by @iholder101) |
| **Dialogs stay on their own workspace** | Dictation Pad, Settings and other standalone windows no longer bind to other windows on Wayland, so opening Settings no longer summons or minimizes the Dictation Pad with it (#1024; fixes #984 reported by @hopsayer) |
| **Ctrl+mouse-wheel zoom works** | Ctrl key events pass through for applications again, so holding Ctrl and scrolling zooms in Firefox and elsewhere even with a double-tap Ctrl shortcut configured (#967 by @loject; fixes #950 reported by @loject) |
| **Upgrade mode in the installer** | `./install.sh` on an existing install offers and defaults to an upgrade that keeps your configuration and models, and `./install.sh --upgrade` runs it non-interactively (#1026; fixes #1009 reported by @plaffitt) |

### Also in v0.18.5

- Paste-injection copies are marked sensitive (`wl-copy --sensitive`) on wl-clipboard 2.2+, so clipboard managers that honor the hint, such as Klipper or cliphist, keep dictated text out of their history; the xclip/xsel path and older wl-copy have no such hint (#956 by @Nosion)
- The tray prefers libayatana-appindicator-glib when it is available (#1019; fixes #1013 reported by @jo-so)
- The Wayland fallback overlay stays alive when the dock closes it, covering the leftover case from 0.18.4 (#1017; fixes #1010 reported by @lavoie005)
- uv pins are aligned across CI and the release workflows (#1016 by @sesav), release build tooling is pinned with a reproducible-wheel gate (#945 by @sesav), the supported uv range is widened with Fedora build hardening (#952 by @sesav), and Dependabot bumps cover pip and GitHub Actions plus flagged vulnerable dependencies (#1014, #1015, #1008 by @Mr-Sunglasses)

See the [full changelog](https://git.bongbetic.com/xavierk/voculinux/releases/tag/v0.18.5).

---

## What's New in v0.18.4

0.18.4 is a **patch** on the stable line. Default engine is still whisper.cpp. It closes the reports filed after 0.18.3: long dictations keep all their audio, model downloads are always your decision, Test Dictation and the Dictation Pad stop competing for text, and GPU and thread selection match what whisper.cpp actually runs on.

### 0.18.4 highlights

| Fix | Description |
|---------|-------------|
| **Long dictations kept whole** | Push-to-talk recordings past the buffer cap are queued for transcription in segments, so a 44.1 kHz mic no longer loses everything after about two minutes (#976 by @dishanood; fixes #975 reported by @dishanood) |
| **Explicit model downloads** | Settings never schedules a download on its own. Every download asks first with its size, Cancel cancels the whole change without freezing the dialog, picking a model keeps your language settings, and [docs/MODEL_SELECTION.md](MODEL_SELECTION.md) has a hardware to model guide (#1006; fixes #979, #980, #981, #982, #985, #988 reported by @hopsayer) |
| **Test Dictation and the Dictation Pad** | Test Dictation always shows its own result, and "Dictate into Dictation Pad" now lives in Settings > Dictation next to an Open Dictation Pad button (#1005; fixes #983, #986 reported by @hopsayer) |
| **GPU picker and CPU threads** | The GPU list follows ggml's own device enumeration, the log reports the backend actually in use, and CPU runs size threads from your core count instead of 2 (#1001; fixes #987, #990 reported by @hopsayer) |
| **Cancel a transcription** | Press the recording hotkey again, use the tray item, or click Cancel in Test Dictation to drop a slow transcription. Nothing is typed or added to history (#1004; fixes #991 reported by @hopsayer) |
| **Slovenian** | `sl` is a supported language in Settings and in language shortcuts (#995; fixes #970 reported by @samodadela) |

### Also in v0.18.4

- Keyboards whose evdev node also reports pointer axes, such as the Logitech K860, are monitored again without being grabbed, so `right_alt` and other hotkeys fire (#978 by @ikarius; fixes #953 reported by @dherdt)
- RemoteDesktop portal requests each get a unique handle token, which fixes the missing GNOME consent dialog and the 15 second timeout (#996; fixes #993 reported by @hal2100)
- XWayland apps that are not connected to IBus, such as SoftMaker TextMaker, receive text through xdotool (#1003; fixes #972 reported by @kdseattlewa)
- On GNOME Wayland the dictation overlay is mapped at startup, so the first recording no longer steals focus from the app you are typing into (#998; fixes #992 reported by @lavoie005)
- `install.sh --engine=parakeet` installs a working Parakeet engine again (#1000; fixes #973 reported by @Ma6nus)
- Update checks follow the build channel, so stable installs only hear about stable releases, and Settings can turn them off (#999; fixes #977 reported by @hopsayer)
- Faster Settings, typing and dictation, plus mypy as a CI gate (#968)

---

## What's New in v0.18.3

0.18.3 is a **patch** on the stable line. Default engine is still whisper.cpp. This one is all fixes: "new line" and "new paragraph" finally insert real line breaks, hotkeys come back on GNOME Shell X11, and release builds stop crashing on older CPUs.

### 0.18.3 highlights

| Fix | Description |
|---------|-------------|
| **Real line breaks** | "new line" and "new paragraph" insert an actual newline standalone or mid-utterance, and the stray punctuation they used to leave behind is gone (#958, #962; fixes #955 reported by @rastarr) |
| **GNOME X11 hotkeys** | Push-to-talk and toggle shortcuts work on gnome-shell X11 sessions again; the app now prefers the evdev backend there (#963; fixes #954 reported by @jepst) |
| **Layout restore on quit** | Quitting no longer collapses a multi-layout keyboard map (us,de staying us,de) because the IBus engine restore is skipped when a real engine is already active (#963) |
| **Older CPUs unblocked** | Release builds now compile the bundled pywhispercpp at the true x86-64 baseline, ending the SIGILL crash on pre-Haswell processors (#959; fixes #957 reported by @bucketsofmail) |

### Also in v0.18.3

- Transcribe File's modal model-download dialog now closes itself once TinyDiarize lands, releasing its grab so the transcript's Copy and Save As buttons are clickable (#965; found during release e2e testing)

---

## What's New in v0.18.2

0.18.2 is a **patch** on the stable line. Default engine is still whisper.cpp. The headline is a livelier dictation overlay plus a round of fixes: the overlay orb now reacts to live microphone level, and the dictation shortcut can no longer strand a key when you are typing while the app starts.

### 0.18.2 highlights

| Feature | Description |
|---------|-------------|
| **Live mic orb** | Dictation overlay orb pulses and ripples with your real audio level instead of sitting static (#940) |
| **Safer startup grab** | A keyboard with a held key or queued key events is left ungrabbed at startup, so typing during launch cannot stick a key on the desktop (#941) |
| **Older CPU fix** | Bundled pywhispercpp builds no longer use -march=native, so release builds stop crashing on CPUs without the build host's instruction set (#939) |

### Also in v0.18.2

- Installer package map covers cairo headers on openSUSE, Void, Solus and Alpine (#937)
- evdev monitors remapper virtual keyboards ungrabbed so remappers like keyd keep working (#932)

---

## What's New in v0.18.1

0.18.1 is a **patch** on the stable line. Default engine is still whisper.cpp. This one is a reliability release: most of it is fixes, and the new features are plumbing you will notice only when you need it. Settings gains a Proxy page so model downloads and update checks work behind SOCKS5 and HTTP proxies, and an opt-in switch that persists transcripts to disk so a failed injection stops meaning a lost sentence. The tray menu drops Start on Login and About (both already lived in Settings).

### 0.18.1 highlights

| Feature | Description |
|---------|-------------|
| **Proxy settings** | New Settings -> Proxy page: off, system, or manual SOCKS5/HTTP CONNECT with optional auth; covers every model download and the update checker (#909, fixes #655) |
| **Transcript persistence** | Opt-in JSONL history on disk; a transcript survives restarts and failed injections (#907, fixes #758) |
| **Slimmer tray menu** | Start on Login and About removed from the tray; both already live in Settings (#905, fixes #654) |
| **Staged model picker** | Speech Model Advanced pickers no longer start a download per click; changes stage until you confirm Download (#908, fixes #894) |
| **Wayland pad fix** | Dictation Pad no longer ghosts away after sitting idle under a Wayland compositor (#912, fixes #896) |

### Also in v0.18.1

- Startup reuses an on-disk same-size whisper.cpp weight instead of re-resolving or re-downloading it (#923, fixes #916 reported by @blacxsnow)
- Text injection ends option parsing before typed text, so a chunk starting with `-` no longer fails as an unrecognized option and truncates the dictation (#922, fixes #921)
- KDE Plasma 6 Wayland: KWin VirtualKeyboard detection fixed, so dictated letter case stops scrambling (#919, fixes #911 reported by @blackde5ert)
- evdev keyboard discovery tightened: pointer-motion devices, multitouch-only touchpads, and devices without a real keyboard key are no longer grabbed as keyboards, so hotkey enablement stops breaking Logitech mice and Goodix touchpads (#902, #917, #918, fixes #900 reported by @brainygamer, #914 reported by @pylame22, #915 reported by @uncletoxa)
- Test Dictation transcription readout moved above its controls in the sidebar footer (#904, fixes #677 reported by @hopsayer)
- `uninstall.sh` no longer deletes files in the directory you ran it from (#901, fixes #897 reported by @SilverChatte)
- Settings > Advanced Initial Prompt accepts typed input again (clicks were being stolen by the enclosing row), and transcript and log saves no longer write a doubled `.txt.txt` name (#927)

---

## What's New in v0.18.0

0.18.0 is a **minor** on the stable line. Default engine is still whisper.cpp. The headline fix is a grab of the dictation hotkey so the shortcut stops leaking into the app under it. Alongside it is a set of Wayland-native paths: an in-app Dictation Pad that keeps text off the fragile injection routes entirely, PipeWire capture (microphone and system audio), and a RemoteDesktop portal injection backend. Dictation gains per-language shortcuts, a tray history menu, a floating overlay, and audio ducking. Long-requested contributor work lands here too: D-Bus activation for compositor global shortcuts, bilingual language candidates, a postprocessing script hook, a custom dictionary with corrections, and file transcription with speaker labels.

### 0.18 series highlights

| Feature | Description |
|---------|-------------|
| **Dictation Pad** | In-app window that receives dictation; you copy text out by hand, so Wayland injection quirks cannot touch it (#887, fixes #726) |
| **Hotkey suppression** | evdev grabs the dictation shortcut and forwards everything else through a uinput clone; the key no longer types into the focused app (#873, #893, fixes #871) |
| **PipeWire capture** | Mic and system-audio sources through the native PipeWire path, not just PortAudio (#889, #883, fixes #751, #760) |
| **RemoteDesktop injection** | Wayland text injection through the RemoteDesktop portal, no ydotoold needed (#885, fixes #750) |
| **Per-language shortcuts + history** | A dictation shortcut per language, layout-follow while dictating, and recent dictations in the tray menu (#880, #837, #487, fixes #805, #821) |
| **File transcription with speakers** | `--transcribe-file` with TinyDiarize per-speaker labels, plus a transcript viewer (#884, fixes #756) |
| **D-Bus activation** | Opt-in D-Bus methods so compositor global shortcuts (and scripts) can start dictation (#568, fixes #761) |
| **Custom dictionary** | Terms bias plus transcript corrections from one file (#890) |

### Also in v0.18.0

- Postprocessing script hook pipes transcriptions through a user command (#479 by @karottenreibe)
- Bilingual dictation: configurable second-language Whisper candidates, with deferred settings edits (#424 by @juanfradb)
- Optional verified Orukeet model for the Parakeet engine (#840 by @Nathan-Roll1)
- Keep recording while an idle-unloaded model reloads instead of dropping the utterance (#851 by @mre31)
- Floating glowing dictation overlay and lowering of other audio while dictating (#516, #861)
- `config.json` can pin the text-injection backend (#649 by @HashimAbdulaziz, fixes #476 reported by @waldemar-p)
- Alt+Shift and Win+Space layout switching keeps working on GNOME Wayland (#876, fixes #848 reported by @hopsayer)
- Shortcut recorder learns unmapped F19/F24 and XF86-aliased F13-F23 (#844, fixes #843 reported by @bisgardo)
- Vosk refuses to load a model that would exceed available memory instead of tripping the OOM killer (#850 by @AmirF194, fixes #676 reported by @hopsayer)
- IBus guard no longer flips GNOME/X11 to a US layout (#827 by @AmirF194)
- Model downloads cancel cleanly even when the fetch stalls, and the download dialog shows verifying instead of a stalled 100% (#888, #864 by @guilhermefeitosa66, fixes #679, #863)
- Settings: scrollable category list, Dictation Tone grayed while sound effects are off, Test Dictation textbox artifact and missing transcription fixed, speech-model follow-ups (#886, #877, #853, #836, fixes #678, #849, #847, #720, #834)
- Update checker falls back when GitHub API rate-limits instead of erroring (#846, fixes #845 reported by @lmstud)
- Audio: stop passing an explicit index when zero devices enumerate; View Logs closes via the titlebar X (#891, #892)
- Installer installs hash-pinned dependencies and build tools, and is split into sourced modules with a generated distro package map (#856, #862, #872 by @sesav)
- The local gateway can run from Settings → Advanced (podman-first) (#774)
- Dependency advisories cleared and workflow tokens scoped to repository reads (#869, #865, #866 by @Mr-Sunglasses, #870)
- Settings dialog consolidated: guard flags and widget construction in one place each (#878, fixes #793)

---

## What's New in v0.17.0

0.17.0 is a **minor** on the stable line. Default engine is still whisper.cpp. This release adds two optional local engines (Faster Whisper and Parakeet) and a simpler Speech Model page. It also fixes XWayland/layout paste, a KDE xdotool crash, push-to-talk tray redraw, clipped start/stop cues, HDA analog mics that abort when opened below native channel count, native GTK installs that ignored OS dark/light, and English-only models that hid every other language.

### 0.17 series highlights

| Feature | Description |
|---------|-------------|
| **Faster Whisper** | Optional CTranslate2 / INT8 Whisper on CPU (`--engine=faster_whisper`) (#543) |
| **Parakeet** | Optional Parakeet TDT 0.6B via sherpa-onnx; default v3-european bundle (`--engine=parakeet`) (#802) |
| **Speech Model simple setup** | Language + speed/accuracy first; engine/size under Advanced (#801) |
| **First-run language** | Seeds from keyboard layout / locale; saved choice is left alone (#796) |
| **Searchable open picker** | Language list filters while open (#798) |
| **Localized punctuation commands** | it/fr/de/es/pt/nl/pl/ru; English phrases still work (#642) |
| **Bare F-keys** | F1–F24 are valid push-to-talk shortcuts (#815) |

### Also in v0.17.0

- Recommended model is a button that sets size and specialization together, and will not ignore a suitable file already on disk (#797)
- XWayland xdotool fallback pastes via clipboard so missing-layout characters are not garbled (#680)
- Clipboard paste uses the key that types **v** on the current layout (#788)
- Timed-out ydotool paste releases Ctrl (#675)
- Read `WM_CLASS` with `xprop` instead of crashing `xdotool getwindowclassname` on KDE Plasma Wayland (#807)
- Applying an already-downloaded model no longer freezes Settings (#790)
- Selected language picks the whisper.cpp variant (English → `.en`). Picking another language, or auto-detect, switches off `.en` / `.en-q*` instead of hiding the rest of the list (#795, #780, #826)
- Speech Model labels are nouns (Language, Other languages, Speed vs accuracy). Unused downloads is a sibling expander of Advanced, not nested inside it (#826)
- Push-to-talk tray icon turns red on every hold (#809)
- Sink-wake no longer clips start/stop cues (#804)
- Open 3-8 channel HDA analog mics at native layout instead of forcing 2ch/48kHz, so PortAudio no longer aborts after read() (`free(): corrupted unsorted chunks`); capture is downmixed to mono for engines (#829, fixes #813)
- Native GTK installs (`install.sh`) follow OS dark/light via the appearance portal then gsettings; skipped when `GTK_THEME` is set so user overrides stay in charge (#830, fixes #816)
- Parakeet keeps recognizing after a decode error (#803)
- Hash-pinned exports for VOSK, Parakeet, and Faster Whisper extras so `install.sh` no longer installs those unpinned; extras are version-capped (#828)
- `just verify-release` checks the published GitHub Release (#791)
- Distro matrix runs `install.sh` (#810)
- `install.sh` Faster Whisper box is CPU CTranslate2 / INT8, not NVIDIA CUDA. The whisper.cpp GPU step does not install the CUDA toolkit; Vulkan is first, CUDA only if a toolkit is already on the machine (#826)

Faster Whisper and Parakeet are optional extras on `install.sh` / source installs. They are not part of the Void package.

---

## What's New in v0.16.2

0.16.2 is a **stability patch** on the 0.16 series. The feature set is the same as 0.16.x. This release fixes KDE leftover IBus so dictation types into real apps, Wayland and IBus keyboard shortcuts through wtype/ydotool, real BackSpace for "delete that", Fedora/Arch installer glslc packages, nightly version stamping, and release integrity pins. CI now gates the distros the docs promise.

### 0.16 series highlights

| Feature | Description |
|---------|-------------|
| **Update checker** | Settings → About checks stable/nightly; tray shows Update Available for newer GitHub releases (#631, #645) |
| **Right Alt PTT default** | New installs default to hold Right Alt (push-to-talk); existing configs keep their shortcut (#648) |
| **Searchable languages** | Type to filter the Speech Model language list (#672) |
| **Delete unused models** | Remove leftover downloaded speech models from Settings (#671) |
| **AGPL-3.0** | License is AGPL-3.0 (#660) |
| **Family mic icons** | App icon, tray states, and site favicons use the shared Voca family mic (#704) |
| **Tone picker** | Settings → Audio: Lift, Flick, Ember, Step, Voca, Soft, Chirp, Scale, Drop, Glass, Off, plus Preview. New installs default to Voca. Catalog uses family preview WAVs (#707, #708) |
| **Installer** | Justfile, uv lockfiles, distro python3-gi required (no pip sdist of PyGObject). Epic #701 still open (#700, #705, #706) |

### Bug fixes in v0.16.2

- **KDE inject**: skip leftover IBus when it is not the session IM so dictation types into Kate, browsers, and terminals (#753, fixes #752 by @justTravis)
- **Wayland shortcuts**: wtype and ydotool deliver real chords instead of refusing or silently doing nothing (#715 by @eiseleb47)
- **IBus shortcuts**: route X11_IBUS through xdotool and WAYLAND_IBUS through wtype/ydotool (#716 by @eiseleb47)
- **Delete that**: send real BackSpace key events instead of U+0008 text (#714 by @eiseleb47)
- **Installer**: Fedora and Arch need glslc/shaderc packages, not glslang (#763, see #604)
- **Nightly**: stamp `version.py` before `python -m build` so wheel metadata matches the filename (#762 by @sesav)
- **Release integrity**: checksums, signatures, and pinned builders for release artifacts (#759 by @sesav, epic #701 phase 5)
- **CI / docs**: test the distros the docs promise and fix docs drift (#773 by @sesav)
- **Site**: the gateway card is Beta; README logo, badges, and privacy copy (#765, #764)

---

## What's New in v0.16.1

0.16.1 is a **stability patch** on the 0.16 series. The feature set is the same as 0.16.x. This release fixes wrong-model startup, leftover tray state after toggle stop, paste into terminals, and GNOME XWayland layout after inject. The installer now requires Python 3.11 and verifies model downloads.

### 0.16 series highlights

| Feature | Description |
|---------|-------------|
| **Update checker** | Settings → About checks stable/nightly; tray shows Update Available for newer GitHub releases (#631, #645) |
| **Right Alt PTT default** | New installs default to hold Right Alt (push-to-talk); existing configs keep their shortcut (#648) |
| **Searchable languages** | Type to filter the Speech Model language list (#672) |
| **Delete unused models** | Remove leftover downloaded speech models from Settings (#671) |
| **AGPL-3.0** | License is AGPL-3.0 (#660) |
| **Family mic icons** | App icon, tray states, and site favicons use the shared Voca family mic (#704) |
| **Tone picker** | Settings → Audio: Lift, Flick, Ember, Step, Voca, Soft, Chirp, Scale, Drop, Glass, Off, plus Preview. New installs default to Voca. Catalog uses family preview WAVs (#707, #708) |
| **Installer** | Justfile, uv lockfiles, distro python3-gi required (no pip sdist of PyGObject). Epic #701 still open (#700, #705, #706) |

### Bug fixes in v0.16.1

- **Startup**: resolve model size per engine instead of the leftover generic `model_size` key, so a VOSK save does not make whisper.cpp look for a missing medium model (#684 by @kacperpaczos, fixes #681)
- **Settings**: save the model only after the engine loaded it, so a cancelled or failed download does not leave config pointing at a file that is not on disk (#685 by @kacperpaczos)
- **Settings**: unused downloads list sized to its real rows so leftover models are not clipped (#686 by @kacperpaczos, fixes #683)
- **Config**: one ConfigManager for the process so Settings writes are not overwritten by a stale cache (#691 by @kacperpaczos, fixes #689)
- **Tray**: stay idle after leftover transcription on toggle stop (#741)
- **Tray**: the missing-model notification can download the recommended model (#687 by @kacperpaczos)
- **Injection**: Ctrl+Shift+V when pasting into terminals, so Ctrl+V is not treated as verbatim insert (#734)
- **IBus**: sync the XWayland layout from GNOME after scoped inject (#742)
- **Installer**: survive a release without a checksum manifest; stop inventing GPUs (#736 by @sesav)
- **Installer**: leftover engine repair, just `--no-sync`, ggml verify from #713 (#732)
- **Installer**: Python 3.11 floor, uv tooling, verified model downloads (epic #701 phases 2.5 and 5) (#713 by @sesav)
- **Settings / About**: even dropdowns, quieter About, family platform marks (#754)
- **Docs**: screenshots page says v0.16; README family/on-device copy (#749, #733, #735, #737)

---

## What's New in v0.16.0

0.16.0 is a **minor** release on the stable line. It adds an in-app update checker with tray notifications, defaults new installs to hold Right Alt push-to-talk, makes the language picker searchable, lets you delete unused downloaded models, and adds a family dictation tone picker. The license is AGPL-3.0. The installer is hardened (Justfile, uv lockfiles, distro python3-gi). The app icon, tray states, and site favicons use the shared Voca family mic.

### Highlights

| Feature | Description |
|---------|-------------|
| **Update checker** | Settings → About checks stable/nightly; tray shows Update Available for newer GitHub releases (#631, #645) |
| **Right Alt PTT default** | New installs default to hold Right Alt (push-to-talk); existing configs keep their shortcut (#648) |
| **Searchable languages** | Type to filter the Speech Model language list (#672) |
| **Delete unused models** | Remove leftover downloaded speech models from Settings (#671) |
| **AGPL-3.0** | License is AGPL-3.0 (#660) |
| **Family mic icons** | App icon, tray states, and site favicons use the shared Voca family mic (#704) |
| **Tone picker** | Settings → Audio: Lift, Flick, Ember, Step, Voca, Soft, Chirp, Scale, Drop, Glass, Off, plus Preview. New installs default to Voca. Catalog uses family preview WAVs (#707, #708) |
| **Installer** | Justfile, uv lockfiles, distro python3-gi required (no pip sdist of PyGObject). Epic #701 still open (#700, #705, #706) |

### New Features

- **In-app update checker**: Settings → About with stable/nightly channels (#631)
- **Update notifications**: Tray menu Update Available entry and About badge when a newer release exists (#645)
- **Right Alt push-to-talk default**: New installs only; existing configs are unchanged (#648)
- **Searchable language combobox**: Filter the long language list by name or code (#672, fixes #652)
- **Delete unused models**: Remove leftover downloaded speech models from Settings (#671, fixes #650)
- **Disable missing-tray warning**: Settings toggle for desktops that false-positive the tray check (#628, fixes #620)
- **Family mic icons**: App icon, tray states, and site favicons use the shared Voca family mic instead of the old Linux rounded-rect (#704)
- **Family dictation tone picker**: Settings → Audio dropdown (Lift, Flick, Ember, Step, Voca, Soft, Chirp, Scale, Drop, Glass, Off) plus Preview. New installs and unknown saved names default to Voca. A saved catalog id, including Off, is left alone. Enable remains the master mute; Off skips start/stop only (#707)

### Bug Fixes

- **IBus**: Require a restorable engine for scoped injection (#623); restore engine after `register_component` teardown (#643, fixes #558); restore XKB layout after scoped injection on X11 (#665, fixes #664)
- **Injection**: Stop typing `test` during the wtype probe (#627, fixes #622)
- **whisper.cpp**: Skip unsupported `context_params` on pywhispercpp 1.4 (#626, fixes #625); use CUDA device 0 when CUDA-backed (#636); honor bundled GPU libs and skip software Vulkan devices (#674)
- **Audio**: Filter unsafe virtual capture devices (#629, fixes #624); open stereo mics at native channel count (#673, fixes #666); catalog tones are the family preview WAVs, not the synthesized #707 files. `generate_sounds.py` does not clobber catalog ids (#708)
- **Tray / Settings**: Prefer Ayatana AppIndicator on KDE (#621); reuse Settings/Logs windows (#669, fixes #653); separate Close from Test Dictation (#670, fixes #651)
- **Settings**: Test Dictation no longer reports no speech when recognition never started (missing model / auto-pause / live engine out of sync) (#702)
- **Clipboard**: Restore after ydotool clipboard-paste (#588); text-only reads and safer overlapping restore (#646)
- **Installer**: Distro python3-gi is required; pip no longer builds PyGObject from sdist. Unset `XDG_SESSION_TYPE` / `XDG_CURRENT_DESKTOP` no longer crash under `set -u` (#706)
- **Installer / downloads / tests**: Gate `util-linux-extra` to Ubuntu 24.04+ (#635, fixes #526); report failed model downloads (#690); stop the suite from overwriting real `config.json` (#694)

### Docs / maintenance

- Website screenshot refresh for v0.15 (#630)
- Prefer Ayatana AppIndicator in Fedora/Arch packaging hints (#638)
- CUDA device 0 note for dual NVIDIA (#644)
- README badges; gateway rename (#695, #696, #697)
- voculinux.com restyled to the Voca family workbench; Open Graph card uses the flat Tux (#728, #729)
- Website copy drops the stale 100% offline claim (#717)
- Installer hardening, Justfile in place of Makefile, and uv lockfiles with pinned build inputs. Epic #701 remains open (#700, #705 by @sesav)

---

## Older releases

Release notes for v0.15.0 and earlier are not included in this document.

---

## Need help?

- [Installation guide](INSTALL.md)
- [Troubleshooting](TROUBLESHOOTING.md)
- [User guide](USER_GUIDE.md)
- [Support](../SUPPORT.md)
- [Report issues](https://git.bongbetic.com/xavierk/voculinux/issues)