localwhisper 0.1.3

Local, offline push-to-talk and wake-word voice-to-text for Windows — types speech at your cursor.

LocalWhisper 🎙️

Local, offline voice-to-text for Windows. Speak, and it types wherever your cursor is — in any app. No cloud, no account, no data leaves your PC.

Two hotkeys work out of the box:

Trigger	How	Best for
Ctrl + B	Hold and speak, release to insert	quick phrases (walkie-talkie)
Alt + N	Tap to start, tap again to stop	longer dictation (hands free)

There is also an optional "Hey Jarvis" wake word — but it is OFF by default (see Optional wake word below), because an always-on listener can false-trigger and type transcribed audio into whatever window is focused.

Powered by faster-whisper (small.en model) and openWakeWord.

---

Install (easiest — no Python needed)

1. Go to the Releases page and download LocalWhisper-Setup.zip.
2. Extract the ZIP (right-click → Extract All). Keep the folder structure.
3. Double-click Install.bat and approve the Windows admin prompt.

That's it. LocalWhisper installs, starts immediately, and launches automatically every time you log in. A small tray icon shows its state:

• ⚪ grey = idle    🟢 green = recording    🟠 amber = transcribing

Right-click the tray icon to Quit. To remove it later, run Uninstall.bat.

First-launch notes

• Windows SmartScreen may show "Windows protected your PC" — this is expected for a new, unsigned app. Click More info → Run anyway.
• Administrator rights are required because global hotkeys (Ctrl+B / Alt+N) need a system-level keyboard hook. The installer sets this up for you.
• If nothing happens on launch, check the log at %LOCALAPPDATA%\LocalWhisper\localwhisper.log.
• Needs a 64-bit Windows 10/11 PC and a microphone. If the app won't start, install the Microsoft Visual C++ Redistributable (x64).

---

Install via pip (for Python users)

If you already have Python 3.9+:

pip install localwhisper
localwhisper

On first run it downloads the small.en model (~240 MB) and the wake-word models automatically. Run your terminal as Administrator so the global hotkeys work.

---

Optional wake word

The "Hey Jarvis" wake word is disabled by default. It is an always-on microphone listener that types transcribed audio into the focused window, so a false trigger (from a meeting, a video, or background speech) can insert unwanted text. Hotkeys are safer because they only act when you press them.

To enable it (no reinstall needed), create a file at %LOCALAPPDATA%\LocalWhisper\config.json with:

{
  "wakeword_enabled": true,
  "wakeword_threshold": 0.8
}

Then quit and relaunch LocalWhisper (or log out/in). Raise wakeword_threshold toward 0.9 if it still false-fires. To turn it back off, set "wakeword_enabled": false (or delete the file). The same file overrides any other setting in the table below.

Tuning

Every field in localwhisper/config.py can be overridden by the config.json file above — no rebuild required:

Setting	Default	What it does
model_name	small.en	Transcription model. base.en is faster, less accurate.
wakeword_enabled	false	Turn the always-on wake word on/off.
wakeword_model	hey_jarvis	Also: alexa, hey_mycroft, hey_rhasspy.
wakeword_threshold	0.75	Higher = fewer false triggers.
wakeword_patience	3	Consecutive detections required to fire (rejects blips).
endpoint_silence_ms	1500	Silence (ms) that ends a hands-free utterance.
energy_threshold	0.01	Loudness that counts as speech.

---

Build from source / run the tests

git clone https://github.com/Bhargava-R-dev/localwhisper
cd localwhisper
py -m venv .venv
.\.venv\Scripts\python.exe -m pip install -r requirements.txt
.\.venv\Scripts\python.exe -c "import openwakeword.utils as u; u.download_models()"

# run from source
.\.venv\Scripts\python.exe -m localwhisper.app

# run the tests
.\.venv\Scripts\python.exe -m pytest -v

Build the distributable EXE

.\.venv\Scripts\python.exe -m pip install pyinstaller
.\.venv\Scripts\python.exe packaging\build_exe.py

Output lands in dist\LocalWhisper\. See packaging/ for the installer scripts and how the release ZIP is assembled.

---

How it works

One microphone stream feeds a small state machine. When idle it runs the wake-word detector; the two hotkeys post events from a background thread. Any trigger records audio, hands it to the warm whisper model, and pastes the result at the cursor via the clipboard. Everything runs locally on CPU.

License

MIT