Xenith

Ultra-fast, on-device voice assistant for Linux

Xenith is a fully local voice assistant that runs entirely on your hardware with no cloud dependencies. Optimized for Intel Core Ultra processors, it achieves ~1.5-2.5 second response times from wake word to audio output.

Features

• 🎤 Wake Word Detection - Always listening with ultra-low power (~2-3W)
• 🧠 Local LLM - Qwen2.5-1.5B runs entirely on-device
• 🔊 Natural TTS - High-quality Piper neural voices
• ⚡ Streaming Response - Audio starts playing as LLM generates
• 🔒 100% Private - No data leaves your device
• 🎨 Beautiful UI - Animated plasma widget shows voice state

Quick Start

# Install dependencies
make install

# Run Xenith
make run

Say "Hi" to activate, then speak your command!

Performance

Stage	Time
Wake word → Detection	~500ms
STT Processing	~300ms
LLM First Token	~200ms
TTS → Audio	~100ms
Total to First Audio	~1.5s

Architecture

┌─────────────────────────────────────────────────────────────────────┐
│                     XENITH VOICE PIPELINE                           │
├─────────────────────────────────────────────────────────────────────┤
│                                                                     │
│   🎤 Microphone                                                     │
│        │ (continuous audio stream)                                  │
│        ▼                                                            │
│   ┌──────────────────┐                                              │
│   │   Wake Word      │  "Hi" detection                              │
│   │   Detection      │  • Whisper on NPU (~2-3W)                    │
│   │                  │  • 0.5s check interval                       │
│   └────────┬─────────┘                                              │
│            │                                                        │
│            ▼                                                        │
│   ┌──────────────────┐                                              │
│   │   Speech-to-Text │  Whisper STT                                 │
│   │   (STT)          │  • OpenVINO on NPU                           │
│   │                  │  • 0.3s silence threshold                    │
│   └────────┬─────────┘                                              │
│            │ text                                                   │
│            ▼                                                        │
│   ┌──────────────────┐                                              │
│   │   LLM Brain      │  Qwen2.5-1.5B                                │
│   │                  │  • OpenVINO on CPU (fast, ~200ms warmup)     │
│   │                  │  • Token streaming enabled                   │
│   └────────┬─────────┘                                              │
│            │ streaming tokens                                       │
│            ▼                                                        │
│   ┌──────────────────┐                                              │
│   │   Sentence       │  Buffers tokens until sentence complete      │
│   │   Buffer         │  • Min 3 chars, ends on .!?;,:               │
│   └────────┬─────────┘                                              │
│            │ sentences                                              │
│            ▼                                                        │
│   ┌──────────────────┐                                              │
│   │   TTS (Piper)    │  Neural text-to-speech                       │
│   │                  │  • ~100ms per sentence                       │
│   │                  │  • In-memory audio (no file I/O)             │
│   └────────┬─────────┘                                              │
│            │ numpy audio                                            │
│            ▼                                                        │
│   ┌──────────────────┐                                              │
│   │   Audio Player   │  Real-time playback                          │
│   │                  │  • Direct sounddevice output                 │
│   │                  │  • 10ms queue polling                        │
│   └────────┬─────────┘                                              │
│            │                                                        │
│            ▼                                                        │
│   🔊 Speakers                                                       │
│                                                                     │
└─────────────────────────────────────────────────────────────────────┘

Data Flow (Optimized)

All audio data flows in-memory with zero file I/O in the critical path:

Mic → numpy → STT(NPU) → text → LLM(CPU) → tokens → TTS(CPU) → numpy → speakers
      ↑                                                              ↑
      └── in-memory ─────────────────────────────────────────────────┘

Hardware Requirements

Minimum

• Intel Core Ultra (Meteor Lake) or newer
• 8GB RAM
• 5GB disk space

Recommended

• Intel Core Ultra 7/9
• 16GB RAM
• Intel Arc GPU (optional, for larger models)

Configuration

Edit config/config.yaml:

llm:
  # CPU recommended for fast response (~200ms warmup)
  # NPU is low power but slow (~2.3s warmup per query)
  device: "CPU"
  model: "qwen2.5-1.5b"

audio:
  stt:
    device: "auto"  # NPU → Intel GPU → CPU
    model: "base"
  tts:
    voice: "EN-Default"  # Ryan male voice (high quality)

Device Trade-offs

Device	LLM Warmup	Power	Best For
CPU	~200ms	~15-30W	Fast response (recommended)
NPU	~2,300ms	~3-5W	Battery life
GPU	~300ms	~30-50W	Larger models

Project Structure

src/
├── app.py                  # Main GTK application
├── main.py                 # Entry point
├── widgets/
│   └── plasma_widget.py    # Animated voice indicator
└── audio/
    ├── voice_input.py      # Wake word & STT handling
    ├── streaming_pipeline.py  # LLM + TTS streaming
    ├── pipeline_metrics.py    # Performance tracking
    ├── stt_backends/       # Speech-to-Text
    │   ├── openvino_backend.py
    │   └── whisper_backend.py
    ├── llm_backends/       # Language Models
    │   └── openvino_backend.py
    └── tts_backends/       # Text-to-Speech
        ├── piper_backend.py
        └── melotts_backend.py

Testing

# Test STT backends
python test_stt_backends.py

# Test TTS backends
python test_tts_backends.py

# Test LLM backends
python test_llm_backends.py

# Test full pipeline
python test_streaming_pipeline.py

Documentation

• Voice Pipeline Architecture
• Intel NPU Setup Guide
• STT Backends Reference
• TTS Backends Reference
• LLM Backends Reference
• Scripts Reference

Performance Tuning

For Fastest Response (~1.5s)

llm:
  device: "CPU"  # 12x faster than NPU
audio:
  tts:
    voice: "EN-Fast"  # Medium quality, faster synthesis

For Lowest Power (~3-5W active)

llm:
  device: "NPU"  # Slower but efficient
audio:
  stt:
    device: "NPU"

License

MIT License - See LICENSE file for details.