piroTalk - Voice-to-Type for Linux

Hold Pause, speak, release — transcribed text appears wherever your cursor is. The hold-to-talk key is configurable (see Changing the hotkey).

Runs entirely on local hardware. Two backends available:

Mode	Model	Hardware	Quality
iGPU (default)	CohereASR 2B (Parakeet + Cohere decoder)	Intel integrated GPU	High accuracy, 14 languages
NPU	Whisper	Intel NPU (Arrow Lake / Meteor Lake)	Fast, 100 languages

---

Requirements

• OS: Ubuntu 24.04 (Wayland or X11)
• Common: xdotool, wtype, libportaudio2, gir1.2-ayatanaappindicator3-0.1

iGPU mode additionally requires:

• Intel integrated GPU supported by OpenVINO
• OpenVINO GPU driver (intel_gpu via Mesa or oneAPI)
• ~1.4 GB iGPU VRAM for the INT4 quantized models
• ~8 GB free disk space to download the model from Hugging Face

NPU mode additionally requires:

• Intel NPU (Arrow Lake / Lunar Lake / Meteor Lake)
• NPU driver installed (intel_vpu kernel module, /dev/accel/accel0 present)

---

Quick Start — iGPU (default)

1. Clone and install dependencies

git clone https://github.com/lucaspirola/piroTalk.git
cd mic
./install.sh

This creates a .venv and installs all Python dependencies into it.

2. Download the model

pip install huggingface-hub
huggingface-cli download CohereLabs/cohere-transcribe-03-2026 \
    --local-dir ~/models/cohere-transcribe-03-2026

Then update MODEL_ID in transcriber_igpu.py to match your path:

MODEL_ID = "/home/yourname/models/cohere-transcribe-03-2026"

3. Quantize to INT4 (one-time, ~15 min)

The model is 2B parameters (~8 GB in FP32). We export it to OpenVINO and compress weights to INT4, reducing iGPU memory usage from ~3.7 GB to ~1.4 GB:

./quantize.py

This exports and quantizes both the encoder (~1.1 GB) and decoder (~87 MB) to ~/.cache/cohere-asr/openvino/. Only needs to run once.

To force a re-export (e.g. after changing MODEL_ID):

./quantize.py --force

4. Run

./pirotalk_igpu.py

Wait for the tray icon to turn green, then hold Pause to dictate.

---

Quick Start — NPU

./install.sh --npu
python3 pirotalk_npu.py

No quantization step needed — Whisper exports automatically on first run (~30 seconds for the base model).

---

How it works

Hold Pause → record from microphone
                  ↓
Release Pause → encoder runs on iGPU or NPU (~0.5s)
                  ↓
              decoder runs on iGPU or CPU (~0.2s)
                  ↓
              text pasted into the window that was focused
              when you started holding the key

You can freely switch windows while waiting — the text will always land in the app you were in when you pressed Pause.

A system tray icon shows the current state:

• Grey — loading model
• Green — ready
• Red — recording
• Blue — encoder running
• Orange — decoder running

---

Configuration

Edit constants at the top of each file:

File	Constant	Default	Description
mic.py	HOTKEY_CODE	119 (Pause)	evdev keycode for hold-to-talk
mic.py	MIC_DEVICE	FHD Camera Microphone	Microphone name substring to match
mic.py	AUTO_ENTER	False	Press Enter automatically after each dictation
transcriber_igpu.py	MODEL_ID	(your path)	Path to CohereASR model
transcriber_igpu.py	LANGUAGE	en	Target language code
transcriber_igpu.py	OV_DEVICE	GPU	OpenVINO device (GPU, CPU)
transcriber_npu.py	WHISPER_MODEL	base	Whisper model size (tiny–large)
transcriber_npu.py	LANGUAGE	en	Target language (None = auto-detect)

Changing the hotkey

Run this to find the evdev keycode for any key on your keyboard:

python3 -c "
import evdev, selectors
devs = [evdev.InputDevice(p) for p in evdev.list_devices()]
sel = selectors.DefaultSelector()
for d in devs: sel.register(d, selectors.EVENT_READ)
print('Press the key you want to use...')
for key, _ in sel.select():
    for e in key.data.read():
        if e.type == evdev.ecodes.EV_KEY and e.value == 1:
            print(f'HOTKEY_CODE = {e.code}  # {evdev.ecodes.KEY[e.code]}')
            raise SystemExit
"

Then set HOTKEY_CODE and HOTKEY_NAME in mic.py to match.

---

Architecture

mic.py              Shared orchestrator: GTK tray icon, state machine,
                    audio capture, evdev key listener.
                    VoiceTypeDaemon(transcriber_class) accepts any
                    transcriber with load() / encode() / decode().

pirotalk_igpu.py         Launcher: iGPU mode (CohereASR via OpenVINO)
pirotalk_npu.py          Launcher: NPU mode (Whisper via OpenVINO)

transcriber_igpu.py CohereASR: Parakeet encoder on iGPU, Cohere decoder
                    on iGPU. INT4 weight quantization via NNCF.
transcriber_npu.py  Whisper: encoder on Intel NPU, decoder on CPU.

typer.py            Text injection: sets GTK clipboard, pastes via
                    xdotool (Ctrl+Shift+V for terminals, Ctrl+V elsewhere).

quantize.py         One-time export script: downloads model, exports
                    encoder + decoder to OpenVINO INT4.

---

Autostart

To have piroTalk start automatically on login:

./autostart.sh           # iGPU (default)
./autostart.sh --npu     # NPU
./autostart.sh --remove  # disable autostart

---

Logs

~/.local/share/mic/pirotalk_igpu.log
~/.local/share/mic/pirotalk_npu.log

---

License

MIT