Voice to text. Fully local. Nothing leaves your machine.
Hold a key, speak, release. Your words appear wherever your cursor is — in any app, any text box, anywhere on Windows.
No account. No internet after setup. No one listening.
Every other voice typing tool sends your audio to a server somewhere.
| Tool | Where your voice goes |
|---|---|
| Google Dictation | Google's servers |
| OpenAI Whisper API | OpenAI's servers |
| Wispr Flow | Their servers |
| Dragon | Nuance/Microsoft |
| say-it | Your machine. Only your machine. |
This matters if you're a doctor, lawyer, journalist, or anyone who speaks things that shouldn't leave the room.
- Hold Right Alt
- Speak
- Release — text appears where your cursor is
That's it. Works in Chrome, VS Code, Notepad, WhatsApp Web, anywhere.
- 100% local — Whisper runs on your CPU or GPU, nothing is sent anywhere
- Works offline — after the one-time model download, no internet needed ever
- Any app — pastes into whatever window you're working in
- Language modes — switch between Auto, Hindi, or translate-to-English from the tray
- History — right-click the tray icon to see past transcriptions
- Lightweight — lives in your system tray, uses no resources when idle
- Windows 10 or 11 (64-bit)
- Microphone
- Internet connection (first run only, to download the AI model)
- Download
say-it.exefrom Releases - Run it — Windows may show a security warning, click More info → Run anyway
- The tray icon appears (green dot)
- First run downloads the AI model (~450 MB) — wait about 2 minutes
- Done. Hold Right Alt and start speaking
The security warning appears because the app isn't signed with a paid certificate yet. The code is fully open source — you can verify exactly what it does.
Right-click the tray icon to switch:
| Mode | What it does |
|---|---|
| Auto (default) | Detects your language, transcribes as-is |
| Hindi → Hindi | Forces Hindi, outputs Devanagari script |
| Any language → English | Speak anything, get English text |
# Clone the repo
git clone https://github.com/YOUR_USERNAME/say-it
cd say-it
# Install dependencies (Windows)
install.bat
# Run directly
run.bat
# Build .exe
build.batRequires Windows Python 3.10+.
Settings are stored at %APPDATA%\say-it\config.json:
{
"model": "small",
"language": "auto",
"task": "transcribe",
"hotkey": "right alt",
"sample_rate": 16000,
"device": null
}Models (trade-off between speed and accuracy):
| Model | Size | Speed (CPU) | Accuracy |
|---|---|---|---|
tiny |
75 MB | ~1s | Basic |
base |
145 MB | ~1-2s | Okay |
small |
450 MB | ~2-3s | Good (default) |
medium |
1.5 GB | ~5-10s | Better |
large-v3 |
3 GB | ~15-20s | Best |
GPU (NVIDIA) is auto-detected — if you have one, transcription is near-instant.
- Audio is processed entirely on your device by OpenAI Whisper via faster-whisper
- No telemetry, no analytics, no network calls after model download
- No account required
- History is stored locally in SQLite at
%APPDATA%\say-it\history.db - Source code is fully open — read every line
- Transcription: faster-whisper (Whisper on CPU/GPU)
- Hotkey: keyboard
- Tray: pystray
- Audio: sounddevice
- Paste: Win32 clipboard API (raw ctypes)
- Installer with auto-start on Windows boot
- Microsoft Store release (zero install friction)
- Custom hotkey from tray settings
- Mac support
- Punctuation mode
MIT — do whatever you want with it.