VoiceToClipboard

Made for everything - but built with vibe coders in mind. 🎙️→📋

A lightweight, always-on-top Windows widget that turns microphone speech into clipboard text.

Note

Project Status: 🟢 Official Release - v5.0.0 This status must not be changed by AI unless explicitly instructed by the repository owner.

Highlights

• Global recording shortcut, customizable in Settings.
• Click-to-record, click again to submit, or press Esc to cancel.
• Gemini cloud transcription with automatic 5-model fallback (rate-limited models are remembered and skipped until the next day).
• Six local offline models - automatic language detection with no language setting (Tiny is English-focused; the rest are multilingual/bilingual):
• Tiny · Zipformer Transducer INT8 (~293 MB) and Mini · FastConformer Transducer 10 languages (~102 MB) - tiny footprint, instant low-RAM wake.
• Chinese + English (Light) · SenseVoice (~158 MB, ~400 MB RAM) and Chinese + English (Big) · FireRedASR2 (~496 MB, ~1.1 GB RAM) - mainly Mandarin, plus English.
• Light · Whisper Small INT8 (~610 MB, ~550 MB RAM) - the default pick; 99+ languages, high accuracy.
• Big · Whisper Turbo INT8 (~538 MB, ~950 MB RAM) - max precision, distilled 4-layer decoder.
• Audio capture uses the native Web Audio API and MediaRecorder; no ffmpeg or sox dependency is required.
• Click-through transparent areas, drag-to-move interaction, system tray controls, visual feedback, silence auto-stop, and idle fade.
• Frameless, restyled Settings window with a custom drag bar and minimize/close controls.
• Auto-Stop includes a live mic meter, a Reset button, and a configurable Auto-Calibrate that measures room noise then speech (2–5 s per phase) and picks a robust percentile-based threshold.

Controls

Action	Shortcut / Control
Start recording	Global shortcut or click the mic
Finish and transcribe	Global shortcut or click the mic again
Cancel	Esc or the stop button
Move the widget	Hold the mic and drag
Open settings	Hover the top strip / record button and click the gear, or use the tray
Quit	Hover the widget and click X, or use the tray

Install

Recommended: download the executable

Grab the ready-to-run build from the Releases page:

• VoiceToClipboard.exe - unpacked app, no installer needed; just download and run.
• VoiceToClipboard-Setup-*.exe - NSIS installer that adds a Start-menu shortcut.

No Node.js, no npm, no dependencies. The app is a single Windows executable (x64, Windows 10/11).

The API key (only needed for Gemini cloud transcription) can be supplied through the GEMINI_API_KEY environment variable or saved through Settings; environment variables take precedence. Keys are never sent to the renderer or printed intentionally.

Alternative: run from source

For developers or the curious - npm is not required to use the app, but if you want to hack on it:

git clone https://github.com/TruftedBug89/VoiceToClipboard.git
cd VoiceToClipboard
npm install
npm start

Requires Node.js 18+ and Windows 10/11.

Local offline models

Select Offline Models in Settings, pick one of the six models from the Offline Local Models list, then download the verified model package. The Settings panel shows a model card with the exact download size, RAM estimate, license, and installation state, plus a single Download & Activate button with inline progress (no separate modal). Installed models can be removed from the same card. Model data is stored outside the installed application under the canonical Electron user-data directory's models folder and is not committed or bundled into releases.

Downloads: .tar.bz2 model packages are decompressed with unbzip2-stream before extraction - the npm tar package alone cannot decode bzip2 archives (older builds failed with invalid base256 encoding). Extraction keeps only the files the app actually loads (INT8 weights, tokens), skipping the fp32 duplicates bundled in Whisper-style archives. If the GitHub release download fails on your network (TLS resets), every model except mini-multilingual automatically falls back to per-file downloads from its HuggingFace mirror.

Model	Languages (auto)	Download	RAM while loaded
Tiny · Zipformer Transducer INT8	English	~293 MB	~180 MB
Mini · FastConformer Transducer	10 languages (EN/DE/ES/FR/IT/PL/RU/UK/HR/BE)	~102 MB	~270 MB
Chinese + English (Light) · SenseVoice	Mandarin, English, Cantonese, Japanese, Korean	~158 MB	~400 MB
Light · Whisper Small INT8	99+ languages	~610 MB	~550 MB
Big · Whisper Turbo INT8	99+ languages	~538 MB	~950 MB
Chinese + English (Big) · FireRedASR2	Mandarin + English (code-switching)	~496 MB	~1.1 GB

All six models detect the language automatically - there is no language setting. The RAM figures are conservative worst-case estimates (peak during transcription plus headroom), so a model that measures ~400 MB at rest is listed around ~550 MB. Model packages carry their own licenses; review the source and license information before redistributing them.

Every model above was downloaded through the real install pipeline and verified live on Windows x64 (English/Spanish samples; Mandarin samples for the Chinese+English pair) - all six transcribe correctly.

Gemini remains the recommended cloud option when a local model is unavailable or when broader language coverage is needed.

Building from source

Build an unpacked app:

npm run pack

Build the Windows installer:

npm run build

Both commands write to the single canonical dist/ directory as configured in package.json. Use npm run clean:dist to remove only that generated output. Generated installers, blockmaps, model archives, logs, and unpacked build output belong in GitHub Releases or local ignored directories, not in Git.

Regenerate the icon with:

npm run icon

Development checks

npm test
npm run check
npm run check:i18n
npm audit --omit=dev

The tests cover model-key selection, legacy config migration, PCM/WAV validation, registry integrity, model-cache path safety, and secret redaction. Full model validation additionally requires Windows x64 and downloading the relevant model archives.

The local inference layer uses a common main-process service and normalized IPC contract. Every backend (Moonshine, NeMo transducer, SenseVoice, Whisper, FireRedASR CTC) has its own adapter branch because their model formats and decoder APIs differ.

License

The application source is distributed under the MIT License. See LICENSE. Model licenses remain those of their respective publishers.