Capture
F9 records the selected microphone with VAD and a visible floating pill.
WINDOWS · LOCAL FASTER-WHISPER · OPEN SOURCE
Wispr Flow magic. 100% offline. $0/month.
Hold F9, speak messily, release, and get clean text in the app already focused. Audio, raw transcripts, cleaned text, glossary, and history stay on your PC.
199tests passed
42/42cleanup evals
0%soak RSS growth
$0monthly software fee
🎙Listening…F9 · 00:07 · LOCAL
SYS · 01
Speech becomes text without becoming cloud data.
Model downloads are explicit setup actions. The dictation path itself makes zero network calls.
Transcribe
Local faster-whisper runs base, small, or medium.
Rewrite safely
Rules remove fillers, repeat phrases, and spoken punctuation.
Match context
Email, chat, notes, and code-comment profiles adjust tone.
Type locally
Clean text enters the focused app; raw and clean stay in SQLite.
SYS · 02
Install the private dictation path.
Ask your coding agent
Read https://voicetray-offline.vercel.app/llms.txt and https://github.com/tanzir71/voice_tray. Create an isolated Python environment, run the 199-test suite, 42-case eval, and synthetic soak, build the one-folder Windows app, then verify microphone, F9 hold/lock, history, and safe insertion. Do not add cloud STT, telemetry, or a dictation network fallback.
From source
pip install -r requirements.txt
python -B -m pytest tests\ -q
python -B -m voicetray.eval
pythonw -m voicetray
REQ · Windows · Python 3.11+ · microphoneSYS · 03
Dictation that respects the desktop.
Hold + lock modes
Hold F9 for push-to-talk or tap for hands-free lock; F10 records to History only.
voicetray/hotkeys.pyGuarded cleanup
Deterministic rules first; optional local GGUF edits are validated and fall back when risky.
dictation/pipeline.pyRaw + clean history
Local SQLite lets you compare, copy, reinsert, and add glossary words.
voicetray/history.pyPer-app tone
General, email, chat, notes, and conservative code-comment profiles.
app_profiles.jsonGlossary + snippets
Personal names and reusable expansions stay local and deterministic.
glossary.json · snippets.txtSafe foreground insertion
If focus changes unexpectedly, VoiceTray preserves History instead of typing into the wrong app.
insert/inserter.pySYS · 04
Privacy versus cloud polish.
| Option | Price snapshot | Dictation path | Tradeoff |
|---|---|---|---|
| VoiceTray | Free / $0 month | 100% local | Windows; model-size speed/accuracy tradeoff |
| Wispr Flow | $15/mo ($12 annual) | Cloud STT | Polished context; free tier 2k words/week snapshot |
| Typeless | ~$12/mo | Cloud | Strong multilingual convenience |
| Superwhisper | $9.99/mo tier | Local/cloud options | macOS-focused snapshot |
| Windows dictation | Free | OS service | Less cleanup and customization |
Pricing/product snapshot: July 2026. Verify current vendor pricing, limits, platforms, and privacy terms.
Windows only
Global hotkeys, focus tracking, insertion, tray behavior, and packaged acceptance target Windows.
Local accuracy tradeoffs
Base is fastest; small improves accuracy; medium is the best CPU-viable option but slower and larger.
No command mode in v1
VoiceTray inserts cleaned dictation. It does not execute voice commands.
SYS · 05
Before you dictate.
Is audio uploaded?
Never during dictation. Audio, STT, cleanup, history, and insertion are local. Model download is an explicit setup action.
How accurate is it?
Accuracy depends on microphone, language, model size, and speech. Medium is strongest locally; cloud giants may still outperform it.
What languages work?
Whisper is multilingual and VoiceTray can use an explicit language or auto detection.
Does the web demo hear me?
No. It uses canned real-style cleanup examples and requests no microphone permission.