Textify

Textify desktop checks

These checks distinguish actual desktop behavior from automated fixture tests. Use a fresh preview installer. The preview.23 Mac installer is Developer ID signed and notarized; Windows and Linux installers are unsigned. Record the app version, OS version, desktop environment (on Linux), microphone, and pass/fail for each row.

The owner confirmed macOS microphone and text insertion testing. Windows has one physical-PC result, recorded below; Linux remains untested on physical hardware. Final signed Mac fresh-install and update verification on a second Mac remains pending. Automated Linux checks use Xvfb and do not establish GNOME/KDE Wayland behavior.

Windows quick start

1. Download the Windows artifact ZIP from the successful Electron workflow run, extract it, and run Textify-…-win-x64.exe. The preview is unsigned.
2. Open Textify, choose Models, and download Whisper small.en or import the exact signed model file. Allow microphone access when starting the first recording.
3. Hold Textify’s microphone button, speak a sentence, release, choose Copy, and paste into Notepad. Confirm the sentence appears once.
4. Enable the global trigger. Focus Notepad, hold Right Control, speak, and release. Confirm automatic insertion appears once in Notepad.

Core behavior

Check	Expected result	Result
Short accidental trigger tap	No text, no lingering recording	Pending
Silence	No invented text	Pending
Hold, speak, release	One completed result	Pending
Escape during recording	Capture ends, no insertion	Pending
Microphone button, then Copy	Clipboard changes only after Copy	Pending
Close settings window	Tray utility remains; reopening works	Pending
macOS Cmd+Q while idle	Main process and helpers exit; next launch opens normally	Passed locally on 2026-09-22 after native quit timing fix
Quit during recording or recognition	Mic stops; no delayed insertion	Pending
Five-minute cap	Recording stops and produces at most one result	Pending

Mac automatic updates

Install the previous updater-enabled release in /Applications, then publish the new release with its ZIP, blockmap and latest-mac.yml.

Check	Expected result	Result
Launch the previous release	Notification and Settings notice name the new version	Pending
Choose Update	Progress reaches ready; the app keeps working	Pending
Quit and reopen Textify	New version runs; Accessibility, microphone and settings are kept	Pending
Turn off automatic checks, relaunch	No update notice or request to GitHub	Pending

Windows and Linux automatic updates

Install the previous updater-enabled release, then publish the new release with latest.yml, the Windows installer’s blockmap and latest-linux.yml.

Check	Expected result	Result
Windows: choose Update, then quit	The installer runs silently; the new version starts next time with settings kept	Pending
AppImage: choose Update, then quit	The AppImage file is replaced and the new version starts next time	Pending
.deb: choose Update, then quit	An administrator password prompt appears; the new version starts next time	Pending

Permissions, focus, and clipboard

Use test text only. For focus checks, have Notepad and a browser test field open.

Check	Expected result	Result
Deny microphone permission	Actionable failure; no stuck recording	Pending
Unplug selected microphone during capture	Recording stops; retry can use a newly selected device	Pending
Select a missing microphone	No silent switch to another microphone	Pending
Change foreground app before recognition finishes	No automatic paste into the newly focused app	Pending
Focus a password field	Global recording/insertion is rejected when the field is identified as secure	Pending
Clipboard initially contains plain text	Previous text restored after automatic paste	Pending
Clipboard initially contains HTML/rich text or an image	Clipboard preserved, or explicit Copy offered if safe restoration is unavailable	Pending
Copy unrelated text while recognition is running	User’s newer clipboard contents are preserved after paste	Pending
Exclude Notepad, then use global trigger there	Microphone does not start	Pending

Models and settings

Check	Expected result	Result
Pause download, quit, reopen, Resume	Valid partial bytes are reused and final checksum is checked	Pending
Import incorrect model bytes	Import rejected; existing model remains usable	Pending
Install two models and choose Use	Selection persists after relaunch	Pending
Remove active model	Dictation becomes unavailable until a model is selected	Pending
Add custom word and replacement pair	Saved across relaunch; replacement remains literal	Pending
Choose Hindi with Whisper large-v3-turbo	Hindi text is preserved without English punctuation-command rewriting	Pending
Enable/disable Launch at login	Registration matches choice on next sign-in	Pending
Change indicator position/scale	Indicator stays within the screen on different monitors	Pending
Confucius live preview exceeds three lines	Older lines move upward smoothly while final text is inserted once on release	Pending

Linux desktop checks

Repeat on GNOME Wayland, KDE Wayland, and X11. Record the desktop version.

• Enable global trigger and accept the desktop shortcut dialog. Hold and release the assigned shortcut; confirm a single recording and a Copy result.
• Deny shortcut consent or use a desktop without GlobalShortcuts. Confirm the microphone button still works and no automatic paste is attempted.
• Cancel from the recording indicator while another app is focused.
• Use Copy, focus the intended app, and paste manually.
• On Wayland, app exclusions must be shown as unavailable because Textify cannot reliably identify the foreground app. On X11, test a configured exclusion.
• Install/relaunch the AppImage or Debian package and recheck shortcut consent, tray reopening, and launch-at-login behavior.

Current evidence

On September 25, 2026, the owner confirmed macOS microphone and text insertion testing. Windows and Linux have not been tested on physical hardware. This report does not verify the final signed preview.23 installer or fresh-install and update behavior on a second Mac; those checks remain pending.

The first CI preview passed native compilation, offline fixture transcription, Electron smoke, packaged launch, and installer creation on all three OSes: https://github.com/Player0109/Textify/actions/runs/35684967292

This does not mark any pending physical-desktop row above as passed. Later feature builds must repeat CI and record their commit and artifact versions.

GPU-required preview regression checks

• Windows NVIDIA: GPU name is shown after loading the model; public phrase is recognized on Vulkan with acceptable latency. Repeat each supported model.
• Linux hardware Vulkan: confirm GPU name and recognition; software-only Vulkan and missing drivers must refuse dictation.
• Insufficient GPU memory or unsupported GPU operations show an error; no CPU inference and no automatic model downgrade occur.
• Mac real microphone: hold Right Command in TextEdit, dictate a public phrase, release, and verify insertion. Record any displayed failure reason.

Additional models and native visual design (September 22)

• Native-style dark sidebar, checkpoint browser and inspector reviewed in screenshots.
• Installed app selects and loads Qwen3-ASR 0.6B Q8_0 on Metal through the UI.
• Installed microphone permission check succeeds after upgrade.
• Re-added the updated app’s Accessibility entry; global trigger reports Hold Right Command.
• Public fixture recognition: Qwen 0.6B Q8_0, Qwen 1.7B BF16 (English/Hindi), Parakeet Q8_0, Confucius Q8_0/F16.
• Fixture capture-to-copy pipeline: Qwen 0.6B, Parakeet Q8_0 and Confucius Q8_0.
• Final installed, signed Parakeet worker recognizes the public fixture using the imported model.
• Owner’s physical speech and insertion check with the updated model selection.

New families currently use Metal on Apple Silicon. No MLX/CoreML variants are claimed for this Electron build. Confucius live previews need physical speech verification before public release.

Implementation 87bc05c passed all three OS jobs, including packaged launch and installer checks: https://github.com/Player0109/Textify/actions/runs/35698952276 This automated evidence does not mark the pending physical-desktop checks passed.

Dock visibility correction (September 22)

• Reproduced hidden Dock state caused by fullscreen-overlay process transformation.
• Regression checks: visible after launch, stays visible with settings closed and overlay shown, Dock activation reopens the same settings window.
• Corrected packaged application installed; process activation policy is regular.
• Owner confirmed the new blue T icon is visible in the bottom Dock.
• Qwen 0.6B fixture capture-to-Copy still passes; Right Command reports ready.

Additional models on Windows and Linux (September 27)

• Windows PC with an NVIDIA RTX 2060: on preview.23, model loading ran until it timed out. An older AMD driver’s switchable-graphics Vulkan layer returned an incomplete device list on every call. With DISABLE_LAYER_AMD_SWITCHABLE_GRAPHICS_1=1 set by hand, the worker loaded on the NVIDIA GPU and the owner’s dictation was transcribed. Workers now start with this variable.
• Mac, rebuilt Metal workers: public fixture recognition for Parakeet Q8_0/Q5_K_M, Qwen 0.6B Q8_0/Q5_K_M (English/Hindi), Qwen 1.7B BF16 and Confucius Q8_0/F16/original BF16 (English/Chinese), plus Confucius live preview.
• Mac, Vulkan builds of both additional workers through MoltenVK (a development check, not a shipped configuration): the same models and fixtures produce the same text as Metal, including Confucius live preview.
• Windows NVIDIA: each new family shows the GPU name after loading and recognizes a public phrase; Confucius shows live text while recording.
• Windows: a new-family model loads from a path containing non-ASCII characters, as in a user profile with a non-ASCII name.
• Windows RTX 2060 (6 GB): Qwen3-ASR 1.7B BF16 and Confucius F16 and BF16 either load and recognize a public phrase, or show the “needs more GPU memory” message, never CPU inference or an endless load.
• Windows: removing, then re-downloading or importing, the active Confucius model succeeds without a file-in-use error.
• Windows icon: the taskbar, title bar, Alt+Tab, notification area and installer show the T mark without the tile, sharp at 100%, 125% and 150% display scaling, on light and dark taskbars.
• Linux hardware Vulkan: untested. CI covers compilation and refusal without a supported GPU only.

Commit 8642e84 passed all three OS jobs, including the Windows build of both additional workers, packaged launch and installer checks: https://github.com/Player0109/Textify/actions/runs/36266929560 The Windows installer’s embedded icon matches the generated icon at all 15 sizes. This automated evidence does not mark the pending physical-desktop checks passed.

This site is open source. Improve this page.