Computer — Local Voice Assistant

A fully offline, private voice assistant for your workstation. No cloud. No API keys. Everything runs on your own hardware.

Pipeline:

Mic → Wake Word ("Computer") → Whisper STT → Ollama LLM → Piper TTS → Speakers

---

Requirements

• Python 3.10+
• NVIDIA GPU with 8+ GB VRAM (48 GB recommended for the 70B model)
• CUDA drivers installed
• A microphone and speakers

Quick Start

1. Clone and install dependencies

git clone https://github.com/JohnTwenty/Computer.git
cd Computer
pip install -r requirements.txt

2. Install Ollama

Windows:

winget install Ollama.Ollama

Linux/macOS:

curl -fsSL https://ollama.com/install.sh | sh

3. Pull a language model

# 70B — best quality, requires ~48 GB VRAM
ollama pull llama3.1:70b

# 8B — good quality, runs on 8 GB VRAM
ollama pull llama3.1:8b

Then create the assistant model with a reduced context window (keeps it 100% on GPU):

ollama create computer-assistant -f Modelfile

4. Download voice and wake word models

python setup_models.py

This downloads:

• Piper TTS voice model → models/
• openWakeWord "Computer" wake word model → models/

5. Run

Make sure Ollama is running (ollama serve on Linux/macOS, auto-starts on Windows), then:

python main.py

Say "Computer" to activate. Say "goodbye" to exit. Say "cancel" to dismiss a false trigger.

---

Configuration

All settings are in config.yaml:

Setting	Default	Description
wake_word.model	computer_v2	Wake word model (filename in models/)
wake_word.threshold	0.98	Detection confidence (0–1). Higher = less sensitive
wake_word.trigger_level	3	Consecutive frames required to activate. Debounces noise spikes
stt.model	large-v3	Whisper model size (tiny / base / small / medium / large-v3)
llm.model	computer-assistant	Ollama model name
audio.silence_threshold	0.02	RMS level below which audio is considered silence
audio.silence_duration	1.5	Seconds of silence before recording stops

Tuning wake word sensitivity

If you get false positives from background noise:

1. Raise threshold (max ~0.99)
2. Raise trigger_level (3–5 works well for most environments)

If it stops responding to your voice:

1. Lower threshold toward 0.85

---

Components

Layer	Library	Notes
Wake word	openWakeWord	ONNX, runs on CPU
Wake word model	fwartner/home-assistant-wakewords-collection	"Computer" v2
Speech-to-text	faster-whisper	GPU accelerated
LLM	Ollama + llama3.1:70b	100% local
Text-to-speech	Piper	en_US-lessac-medium
Audio I/O	sounddevice	Cross-platform

---

Changing the wake word

Other pre-trained options from openWakeWord:

python -c "from openwakeword.utils import download_models; download_models(['hey_jarvis_v0.1'])"

Then set wake_word.model: hey_jarvis_v0.1 in config.yaml.

Available built-ins: alexa_v0.1, hey_jarvis_v0.1, hey_mycroft_v0.1, hey_rhasspy_v0.1

---

Running on a smaller GPU

Edit config.yaml:

stt:
  model: "small"          # instead of large-v3
  compute_type: "int8"    # instead of float16

llm:
  model: "llama3.1:8b"    # pull with: ollama pull llama3.1:8b

And update Modelfile to FROM llama3.1:8b.