🦄 Unicorn Orator - Lightweight TTS with Hardware Acceleration

Efficient text-to-speech that runs on Intel iGPU, freeing your GPU for AI inference

Web Interface | Docker Hub | API Docs

---

🎯 Why Unicorn Orator?

The Problem: Running TTS alongside LLMs fights for GPU resources, slowing down inference and increasing latency.

Our Solution: Unicorn Orator offloads TTS to Intel integrated graphics or AMD NPUs, leaving your discrete GPU free for what it does best - running large language models.

Key Benefits

• 🚀 Free Your GPU: TTS runs on iGPU/NPU, preserving discrete GPU for LLM inference
• ⚡ Resource Efficient: Uses ~15W on iGPU vs 100W+ on discrete GPU
• 🎭 50+ Quality Voices: Kokoro v0.19 with diverse accents and styles
• 🔌 OpenAI Compatible: Drop-in replacement, no code changes needed
• 🐳 Production Ready: Docker image available, battle-tested deployment

🖼️ Web Interface

Clean, intuitive interface with 50+ voices and advanced settings

🚀 Quick Start

Using Docker (Recommended)

# Pull and run the pre-built image
docker run -d --name unicorn-orator \
  -p 8885:8880 \
  -v $(pwd)/kokoro-tts/models:/app/models:ro \
  --device /dev/dri:/dev/dri \
  --group-add video \
  magicunicorn/unicorn-orator:intel-igpu-v1.0

# Visit http://localhost:8885/web for the interface

From Source

git clone https://github.com/Unicorn-Commander/Unicorn-Orator.git
cd Unicorn-Orator
docker-compose up -d

💡 Technical Innovation

Intel iGPU Optimization

We've optimized Kokoro TTS to run efficiently on Intel integrated graphics via OpenVINO:

• Hardware Detection: Automatically detects and uses Intel Xe/Arc iGPUs
• FP16 Inference: Maintains quality while doubling throughput
• Minimal Memory: ~300MB VRAM usage, leaving room for other tasks
• Power Efficient: 10-15W TDP vs 75-350W for discrete GPUs

AMD NPU Support (Experimental)

For Ryzen AI laptops (7040/8040 series), we're developing custom NPU support:

• Custom Runtime: Direct NPU access bypassing standard frameworks
• INT8 Quantization: Optimized models for NPU architecture
• Ultra Low Power: <10W for continuous synthesis

Performance Comparison

Hardware	Power Usage	VRAM	Speed	Purpose
Intel iGPU	15W	300MB	5x realtime	TTS (This Project)
AMD NPU	10W	256MB	4x realtime	TTS (Experimental)
NVIDIA 4090	350W	2GB	20x realtime	Better used for LLMs
CPU (i7)	45W	N/A	2x realtime	Fallback option

📡 API Usage

OpenAI-Compatible Endpoint

import requests

# Works exactly like OpenAI's API
response = requests.post('http://localhost:8885/v1/audio/speech',
    json={
        'text': 'Hello from Unicorn Orator!',
        'voice': 'af_heart',  # 50+ voices available
        'speed': 1.0
    }
)

with open('output.wav', 'wb') as f:
    f.write(response.content)

Available Voices (Selection)

Voice ID	Description	Best For
af_heart	Warm, friendly female	General narration
am_michael	Professional male	News/corporate
bf_emma	British female	Audiobooks
af_bella	Young American female	Social media
bm_george	British male	Documentation

[Full voice list available at /voices endpoint]

🏗️ Architecture

Your System:
┌──────────────────┬──────────────────┐
│   Discrete GPU   │   Intel iGPU     │
│   (RTX/Arc/RX)   │   (Xe Graphics)  │
│                  │                  │
│   Running:       │   Running:       │
│   - LLMs         │   - Unicorn TTS  │
│   - Stable Diff  │   - Video decode │
│   - ML Training  │   - Display      │
└──────────────────┴──────────────────┘
         │                  │
         └──────┬───────────┘
                │
        [High Performance AI]
         Without Competition

🔮 Roadmap

Current Release (v1.0)

• ✅ Intel iGPU support via OpenVINO
• ✅ 50+ Kokoro voices
• ✅ OpenAI API compatibility
• ✅ Docker deployment
• ✅ Web interface

Planned Features

• Real-time streaming
• AMD NPU production support
• Voice cloning (ethical use only)
• SSML support
• Batch processing API
• Kubernetes operator

Future Exploration

• Apple Neural Engine support
• Qualcomm Hexagon DSP
• Edge deployment (Jetson, Pi 5)
• WebGPU browser runtime

🛠️ Building From Source

Prerequisites

• Docker & Docker Compose
• Intel CPU with Xe/Arc graphics (or AMD Ryzen AI)
• 8GB RAM minimum
• Ubuntu 22.04+ or Windows 11 WSL2

Build Steps

# Clone repository
git clone https://github.com/Unicorn-Commander/Unicorn-Orator.git
cd Unicorn-Orator

# Download models (one-time, ~350MB)
./download_models.sh

# Build with hardware detection
./build.sh

# Run
docker-compose up -d

📊 Benchmarks

Testing setup: Intel Core i7-13700K with Intel UHD 770 iGPU

Text Length	Generation Time	Realtime Factor
1 sentence	180ms	5.5x
1 paragraph	950ms	5.2x
1 page	4.2s	5.0x

Realtime factor = audio duration / generation time

🤝 Contributing

We especially welcome contributions for:

• Hardware optimization (OpenVINO, XDNA, CoreML)
• Additional TTS models beyond Kokoro
• Voice training and fine-tuning
• Performance improvements

See CONTRIBUTING.md for guidelines.

🙏 Acknowledgments

• Kokoro TTS - The excellent TTS model we build upon
• OpenVINO Toolkit - Intel's inference optimization framework
• Hugging Face - Model hosting and community

📜 License

MIT License - See LICENSE for details

🏢 UC-1 Pro Ecosystem

Unicorn Orator is part of the UC-1 Pro AI infrastructure suite:

Service	Purpose	Port
Unicorn Orator	Text-to-speech	8885
Unicorn Amanuensis	Speech-to-text	8886
Unicorn vLLM	LLM inference	8000
Open-WebUI	Chat interface	3000

---

Free your GPU. Enhance your AI.

🐳 Docker Hub • 🐛 Issues • 💬 Discussions

Built by Magic Unicorn Unconventional Technology & Stuff Inc.