Changelog

v1.4.1 Current

October 25, 2026

A maintenance release. Updating keeps your settings, history and trial.

Changed

• Free trials are now issued by WhisperType's own server. A trial started on 1.4.0 keeps working until it ends.
• Paid plans need Windows or an Apple Silicon Mac. Intel Macs keep version 1.3.2 with the free trial.

Removed

• The usage-statistics level that sent transcript text. Two levels remain, Off and Anonymous stats (no text); an app that had it on switches to Anonymous stats on update.

v1.4.0

October 11, 2026

WhisperType used to be free. It is now paid, and the app and this site have a new look.

Changed

• WhisperType is now a paid app. 14 days free with an email and no card, then $9 / month. If you were using it for free, this update shows a screen that says so and starts your trial from there.
• The trial is open today. Subscriptions open in a few days. Get an email the day they do.
• New look. The first-run setup was rebuilt; the main window, Settings, History, the floating bar and the lock screen were redrawn. The fonts ship inside the app, so it no longer fetches any from the network.
• No model is reserved for a plan. The "Pro" label on models is gone.

macOS

• Apple Silicon Macs transcribe with whisper.cpp on Metal.
• Every Mac update is signed with the same certificate, so macOS keeps the permissions you granted. Updating from 1.3 asks for the macOS permissions one last time; later updates keep them.
• The microphone is released while the Mac sleeps.
• The audio kept from before the key press went from 1.5 s to 0.5 s, so room chatter stays out of the transcript.
• Each permission request says what the permission is for.
• The trial survives a change of the computer's network name.
• If Accessibility permission is gone, the transcript stays on the clipboard instead of being lost.

Fixed

• A long dictation is no longer lost. The audio is kept while you record, and if transcription fails, the recording goes to History with Retry, even with History turned off. Up to 20 failed recordings (500 MB) are kept.
• A locked app stops dictation in the engine, not only on screen.
• The text you just dictated stays on the clipboard, so Ctrl+V (Cmd+V on macOS) pastes it again in any text box, even when the automatic paste landed nowhere.
• With the language set to Auto, WhisperType chooses English or Spanish from your recording and transcribes in that language, never a translation. For any other language, pick it in Settings.

Known limitations

• The installers are not code-signed. Windows shows SmartScreen ("More info" → "Run anyway"); macOS blocks the first launch until you choose "Open Anyway" in System Settings → Privacy & Security.

v1.3.0 – v1.3.2 macOS

August 2026 — macOS (beta)

• 1.3.2 — the app waits up to 90 seconds for the engine to start instead of 15; slow Intel Macs were restarting it in a loop.
• 1.3.1 — the trial request identifies the app version.
• 1.3.0 — anonymous usage statistics, on by default, with the switch in Settings → Privacy and the detail in the privacy policy. Intel Macs default to the multilingual tiny model instead of tiny.en. Fixed a hang on first launch.

v1.2.0 macOS

July 18, 2026 — macOS (beta)

Mac reliability release: fixes an engine deadlock, a background CPU runaway, and speeds up first-run transcription on Intel Macs.

Fixed

• BUG-041 — the local engine's /transcribe endpoint could deadlock its event loop on a slow transcription, leaving /health and the rest of the API unresponsive until it finished. The blocking call now runs in an awaited executor with a timeout instead of blocking the loop directly.
• BUG-042 — the floating panel's waveform animation kept running on the GPU-helper process even while idle, burning CPU in the background. The render loop is now gated to run only while actively recording or transcribing.
• Intel Macs now default to the smaller, faster tiny.en model on fresh installs instead of base (existing installs keep their current setting).

SHA-256

WhisperType-1.2.0-x64.dmg (installer)

2e2ed006db299885d2f5fd22d2003feef615b4faf648f04170079991f4c61977

WhisperType-1.2.0-mac.zip (auto-update package)

3fd8c7a9554018f528badd00343dee073da66e96fc2fac8664aad59e177c08a8

v1.0.1

June 14, 2026 — Windows + Mac

Unifies the Windows and macOS development lines into one tree and ships the first public binaries for both platforms.

Added

• Native macOS build. New main-window layout (stats strip, dictation card, recent list), visible tray icon, login-item dedup, and an uninstaller. Adhoc-signed .app bundle for a stable CDHash (BUG-020).
• Remote-PC GPU offload (removed). The Mac engine could send transcription to a GPU on another PC, falling back to local on any failure (BUG-021, BUG-022). It is no longer in the app.

Fixed

• BUG-023 — last word clipped on F8-stop. Disabled faster-whisper's 6-step temperature fallback retry that could shift segment boundaries on borderline clips. Per-clip timing log now writes cross-platform.
• Engine startup bind, model auto-load, and single-word clip handling on macOS.

v1.0.0

April 22, 2026 — Commercial Release

Fully self-contained Electron + Python desktop app. No Python install, no CUDA install, no model-download prompt on first run. Ships as a single NSIS installer with an EULA page and a proper uninstall entry.

Core dictation

• Global F8 hotkey to start/stop dictation anywhere in Windows. Survives sleep/wake, re-registers if Windows drops it.
• GPU-first with CPU fallback. Always attempts CUDA/float16, silently falls back to CPU/int8 on any failure. WT_FORCE_CPU=1 skips GPU entirely.
• faster-whisper base model by default — multilingual, bundled, works offline immediately. A larger model downloads at first run for GPU users.
• 100+ languages auto-detected (Since 1.2: Auto (English/Spanish) plus 12 languages). Explicit language override (fix for BUG-009) so Spanish stays Spanish instead of being transcribed as English.
• App-aware text injection — simulated keystrokes into the focused app at the cursor, with a clipboard fallback so Ctrl+V always recovers the transcription.
• Escape cancels recording — audio discarded immediately, nothing transcribed, nothing injected.

AI Polish (opt-in)

• AI Polish providers — optional post-transcription cleanup via local Ollama (default, llama3.2:3b) or OpenAI (opt-in, API key stored in userData). 2-second timeout cap; falls through to raw transcript on error.
• Custom dictionary — force correct spelling of proper nouns, brand names, technical terms. Pushed into the polish prompt on every transcription.
• Text snippets — map short triggers to longer expansions applied after transcription (signatures, boilerplate, code templates).

UI & onboarding

• Floating glass panel — pill-shaped always-on-top mini-UI with a live canvas waveform visualizer. Four states (idle / recording / processing / ready) plus an error-dismiss affordance. Draggable.
• First-run onboarding wizard (8 screens): welcome, privacy choices, mic test, hotkey test, live dictation test, language, stats explainer, finish. Telemetry defaults to OFF.
• Settings panel — change hotkey, language, model, startup behavior, polish provider, and dictionary/snippets from a proper UI. No more config-file editing.
• Transcription history — searchable local log of everything you've dictated. Stays on-disk, never synced anywhere.
• System tray icon — persistent background presence. Right-click for Settings, License, Quit.
• Commercial EULA on the first installer page. User must accept before proceeding.

Reliability

• Single-instance lock — launching a second copy exits cleanly. No duplicate panels, no duplicate tray icons.
• Self-healing watchdog — a per-user scheduled task checks every 5 minutes that the app is alive and relaunches it if not. Registered by the installer, removed on uninstall.
• Auto-start on Windows login (user-toggleable).
• PortAudio hot-plug recovery — re-initialises audio when the default input device disappears (unplugged headset, Bluetooth disconnect, USB re-enumeration) so dictation keeps working on the newly available device without a restart.
• Stall watchdog — detects silently wedged audio streams via a tick counter instead of trusting InputStream.active (which lies on Windows MME/WASAPI under some driver configs) and forces a re-open.
• Pause/resume media automatically during recording via PortAudio peak-meter detection. No false positives on paused Spotify.

Privacy & updates

• 100% local transcription. No audio, no text, no clipboard ever leaves the machine by default; the optional AI Polish (OpenAI) path sends transcript text only — see Privacy Policy.
• In-app update banner — polls GitHub Releases on launch. Transmits only the current version and a generic WhisperType/<version> User-Agent. Clicking opens the release page in the default browser — never silent auto-install.
• Opt-in Sentry crash reporting — disabled by default. When enabled, only stack traces and non-personal diagnostic tags (OS version, GPU, app version) are sent. Before-send scrubbing guarantees no audio, transcribed text, or clipboard content is ever included.

Fixes shipped since first internal builds

• BUG-012 — last 1–2 words clipped from every transcript. Disabled VAD on the live dictation path (VAD is useful for noisy cold-mic clips but harmful for keyed dictation) and added a 100 ms flush before reading the final audio frame. Side benefit: materially fewer hallucinations and better punctuation, because Whisper now sees continuous audio instead of VAD-fragmented segments.
• BUG-011 — legacy tkinter floating panel relaunching from a stale Startup shortcut alongside the Electron app. Shortcut removed, legacy source parked as .legacy-disabled, HKCU Run is now the sole auto-start entry.
• BUG-009 — Spanish (and other non-English) being transcribed as English. Explicit language override on the engine path.
• BUG-008 — Error opening InputStream: Invalid device [-9996] after hot-plug. Device enumeration now re-runs on stream-open failure.
• BUG-007 / BUG-005 — cuBLAS DLL crash on packaged build. cublas64_12.dll and friends bundled correctly; GPU path now works on a clean VM.
• BUG-001 / BUG-010 — recording loop hangs on long audio + "not recording" / "timed out" on long transcribe. Engine state machine rewritten around a single capture dict.

Known limitations

• Installer is unsigned until the EV code-signing cert lands. SmartScreen will show "Windows protected your PC" and require a "More info → Run anyway" click on first install.
• GPU path requires an NVIDIA GPU with driver 525.60+ and ships as a separate larger installer (WhisperType-Setup-1.0.0-gpu.exe).

What's next

Signed installers

Code signing on Windows and macOS, so the first launch needs no "Run anyway" or "Open Anyway".

Coding mode

Code-aware punctuation, PascalCase / snake_case handling, VS Code + Cursor integration hints.

Have a feature request? Reach out or email us.