🎙️ Universal XTTSv2 API (RTX 50-Series Optimized)

A high-performance clone/derivative of lojik-ng/docker-tts-api-ui.

A high-performance, containerised Text-to-Speech API using Coqui XTTSv2. This build is specifically patched to support the NVIDIA Blackwell (RTX 5090/5080) architecture and PyTorch 2.6+ security layers, ensuring near-instant voice cloning on modern hardware.

---

🚀 Features

• Persistent Inference: The model stays resident in VRAM for < 1s generation times.
• Blackwell Support: Custom library patches for RTX 50-series compatibility.
• Zero-Shot Cloning: Clone any voice using a 6-10 second .wav sample.
• Universal Build: Automatically scales down to older hardware (e.g., GTX 1080).
• Dual-Server Architecture: Python Flask engine for AI + Node.js Express for API.

---

🛠️ Prerequisites

Before cloning, ensure the host machine has:

• NVIDIA Drivers: Latest Game Ready or Studio drivers.
• WSL2: Windows Subsystem for Linux (wsl --install).
• Docker Desktop: Configured to use the WSL2 backend.

---

📂 Project Structure

.
├── voices/               # Place reference .wav files here
├── models/               # AI model weights (auto-downloaded)
├── server/
│   ├── index.js          # Node.js API Gateway
│   ├── tts_engine.py     # Persistent Python AI Engine
│   └── public/           # Generated audio files
├── Dockerfile            # Blackwell-ready build
└── entrypoint.sh         # Hardware patching & boot logic

---

⚡ Quick Start

1. Prepare your Voices

Place a clear, 10-second .wav file of the voice you wish to clone into the /voices directory (e.g., hero.wav).

2. Build the Image

docker build -t my-universal-tts .

3. Run the Container

Replace C:\Path\To\Project with your actual local path.

docker run -d -it -p 2902:2902 --gpus all --restart=unless-stopped `
-e TORCH_FORCE_WEIGHTS_ONLY_LOAD=0 `
-v "C:\Path\To\Project:/shared" `
-v "C:\Path\To\Project\models:/root/.local/share/tts" `
-v "/usr/lib/wsl/lib:/usr/lib/wsl/lib:ro" `
--shm-size=8gb `
--name docker-tts-api-ui my-universal-tts

---

📡 API Usage

Generate Voice

POST http://localhost:2902/use-voice

Body (JSON):

{
  "prompt": "The 5090 is officially the king of speech synthesis.",
  "apiKey": "your_key_here",
  "speaker": "hero",
  "language": "en"
}

List Voices

GET http://localhost:2902/list-voices

---

🔧 Hardware Optimization Notes

RTX 5090 / 9800X3D

On this hardware, the first request will take ~15 seconds to load the 2GB model into VRAM. Every subsequent request will be near-instant. Use the --shm-size=8gb flag to prevent memory bottlenecks between the CPU and GPU.

GTX 1080 / Older Cards

This build is "Universal". It will detect older CUDA cores and adjust the kernels accordingly. Ensure you have at least 8GB of VRAM available for stable performance.

---

⚠️ Troubleshooting

1. GPU not "touching" the workload? Ensure you are passing the /usr/lib/wsl/lib volume mount. This is required for Docker to see the Blackwell drivers on Windows.
2. AttributeError / Pickle Errors: These are handled by the entrypoint.sh patches. If they persist, ensure TORCH_FORCE_WEIGHTS_ONLY_LOAD=0 is set in your environment.
3. Robotic Audio: Check your reference .wav file. It should be clean, mono, and roughly 10 seconds long.

---

📄 License

This project is for educational/personal use. Please adhere to the Coqui TTS and Model licenses regarding commercial usage and ethical AI voice cloning.

---

📝 .gitignore Recommendation

node_modules/
models/
public/*.wav
public/*.mp3
logs/*.log
keys.json
.DS_Store