omni-voice

Hold a key → speak → AI agent reasons with tools → speaks back.

Stack: Parakeet TDT (ONNX, local) → Qwen3-6-27b agent → KittenTTS

---

Usage

python QV.py --loop        # hold F9 to record, release to stop, repeats
python QV.py               # one shot
python QV.py --hotkey F12  # different hotkey
python QV.py -q "how much RAM do I have"  # text input

Pipeline

[F9 hold] → [mic] → [Parakeet STT] → [Qwen3 agent + tools] → [KittenTTS] → [speaker]

• STT — Parakeet TDT 0.6B INT8 ONNX, runs fully offline, records at native device rate and resamples to 16kHz
• Agent — Qwen3-6-27b on local GPU via vLLM, agentic loop with 7 tools
• TTS — KittenTTS nano, Bella voice

Agent Tools

bash · read_file · write_file · grep · glob · web_search · web_fetch

Modules

File	Run standalone
stt.py	python stt.py — hold F9, get transcript
kivi.py	python kivi.py "question"
tts.py	python tts.py "say this"
QV.py	python QV.py --loop
test_all.py	python test_all.py

Install

uv pip install onnxruntime sounddevice scipy pynput openai ddgs scrapling kittentts

Parakeet model dir: ~/.local/share/com.pais.handy/models/parakeet-tdt-0.6b-v3-int8/