VocaWin 0.2.0-beta
App version and tag are aligned as 0.2.0-beta / v0.2.0-beta. This GitHub Release is Latest, not a prerelease, so it shows on the repo homepage.
What changed since v0.1.1-beta
- Whisper now runs on transcribe.cpp (Handy's engine), with Vulkan on the GPU VocaWin picks and a CPU fallback. Existing Whisper downloads keep working, so there is nothing to download again.
- New models: Canary 1B v2, Cohere Transcribe, and Voca Hinglish. Parakeet, SenseVoice and Canary use DirectML, with a CPU fallback.
- Long takes are decoded in windows for Canary, Moonshine and GigaAM, so no audio is dropped at the cuts.
- Skip Silence finds speech with Silero VAD. The voice detector downloads with your models and is not bundled in the installer.
- Filler words are removed after the transcript's language is detected, so Portuguese "um" and French "euh" are handled under Auto-detect.
- On-screen pill: a "VocaWin is ready" pill at launch and a live level while you speak. It can optionally show the text as you speak (Settings > Recording overlay, off by default). It never takes focus.
- More ways to dictate: Escape cancels a take. There is a hands-free start/stop shortcut, dictation with a middle or side mouse button, and a paste-last shortcut. Shortcuts are recorded in the keyboard hook.
- Faster starts: the mic stays open for 30 s after a take, and ONNX models stay loaded and begin loading when you press the hotkey.
- Text rules that run on this PC: removes hesitations and stutters, keeps only the correction when you fix yourself, spoken numbers/symbols/emoji (English only), a personal dictionary, and snippets.
- History and stats: search, copy, replay, retry and delete takes. Audio is saved before transcribing. Retention runs from 1 day to forever. Stats show pace, streaks and time saved. All of it stays local.
- Settings are grouped with a search box, and you can export and import a backup. A first-run guide helps you pick a model and try a first dictation.
- Fixes: text insertion is more reliable across apps, Notepad pastes through the clipboard, the launch-at-login path is quoted, and Whisper no longer types [BLANK_AUDIO] for a take with no speech.
Hold Right Alt (or your hotkey) and talk. Text lands where the caret already is. After a model is on disk, audio and speech-to-text stay on this PC.
Still a beta tester build. The NSIS setup is unsigned, so Windows will say the publisher is unknown. Click More info, then Run anyway if you trust this GitHub Release. vocawin.com points here.
NSIS is a current-user setup. MSI is paused for X.Y.Z-beta cuts.
Need today's main instead? The nightly is an unsigned NSIS build from main. Use this tag if you want the named cut.
Open Models and download an engine before the first take. Whisper Tiny is the usual start.
Something broken? Open an issue. Discord is discord.gg/t6muquAJbm.