English | 简体中文 | 繁體中文 | 日本語

---

✨ Features

• 🎨 Beautiful Web UI - Modern interface with real-time audio playback
• 🔌 REST API - Full-featured HTTP endpoints with Swagger docs
• 📡 WebSocket - Real-time bidirectional TTS communication
• 🌊 Streaming - Audio chunks delivered as they generate
• 📦 Batch Processing - Process multiple texts in one request
• 🤖 MCP Server - AI agent integration (Claude, etc.)
• 🌍 Multi-language - English, Chinese, Japanese, Spanish, French, Hindi, Italian, Portuguese
• 🚀 GPU Accelerated - CUDA support with automatic memory management
• 📱 54+ Voices - Wide variety of male and female voices

🚀 Quick Start

docker run -d --name kokoro-tts --gpus all -p 8300:8300 neosun/kokoro-tts:latest

Open http://localhost:8300 in your browser.

📦 Installation

Prerequisites

• Docker 20.10+
• NVIDIA GPU with CUDA support (optional, CPU fallback available)
• nvidia-docker2 (for GPU support)

Docker Run

# With GPU
docker run -d \
  --name kokoro-tts \
  --gpus all \
  -p 8300:8300 \
  -e GPU_IDLE_TIMEOUT=600 \
  --restart unless-stopped \
  neosun/kokoro-tts:latest

# CPU only
docker run -d \
  --name kokoro-tts \
  -p 8300:8300 \
  --restart unless-stopped \
  neosun/kokoro-tts:latest

Docker Compose

services:
  kokoro-tts:
    image: neosun/kokoro-tts:latest
    container_name: kokoro-tts
    ports:
      - "8300:8300"
    environment:
      - GPU_IDLE_TIMEOUT=600
      - KEEP_MODEL_LOADED=true  # Never release model from memory
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    restart: unless-stopped

docker-compose up -d

Verify Installation

# Health check
curl http://localhost:8300/health

# Generate speech
curl -X POST http://localhost:8300/api/tts \
  -H "Content-Type: application/json" \
  -d '{"text":"Hello world","voice":"af_heart"}' \
  -o output.wav

⚙️ Configuration

Variable	Default	Description
PORT	8300	Server port
GPU_IDLE_TIMEOUT	300	Seconds before GPU memory release
KEEP_MODEL_LOADED	false	Never release model from memory (set to true for lowest latency)
NVIDIA_VISIBLE_DEVICES	all	GPU device selection

📖 Usage

Web UI

Tab	Description
Single	Generate single audio file
Stream	Real-time streaming playback
WebSocket	Bidirectional real-time TTS
Batch	Process multiple texts at once

REST API

Generate Speech (WAV)

curl -X POST http://localhost:8300/api/tts \
  -H "Content-Type: application/json" \
  -d '{"text":"Hello world","voice":"af_heart","speed":1.0}' \
  -o output.wav

Generate Speech (Base64)

curl -X POST http://localhost:8300/api/tts/base64 \
  -H "Content-Type: application/json" \
  -d '{"text":"Hello world","voice":"af_heart","speed":1.0}'

Streaming

curl -X POST http://localhost:8300/api/tts/stream \
  -H "Content-Type: application/json" \
  -d '{"text":"Long text here...","voice":"af_heart"}'

Batch Processing

curl -X POST http://localhost:8300/api/tts/batch \
  -H "Content-Type: application/json" \
  -d '{
    "items": [
      {"id":"1","text":"First","voice":"af_heart"},
      {"id":"2","text":"Second","voice":"am_michael"}
    ]
  }'

WebSocket

const ws = new WebSocket('ws://localhost:8300/ws/tts');
ws.onopen = () => {
  ws.send(JSON.stringify({
    text: "Hello world",
    voice: "af_heart",
    speed: 1.0
  }));
};
ws.onmessage = (e) => {
  const data = JSON.parse(e.data);
  if (data.status === 'chunk') {
    // Play audio: data.audio (base64)
  }
};

MCP Integration

{
  "mcpServers": {
    "kokoro-tts": {
      "command": "docker",
      "args": ["exec", "-i", "kokoro-tts", "python", "/app/docker/server.py", "mcp"]
    }
  }
}

🎤 Available Voices

Models

Model	Languages	Voices	Best For
hexgrad/Kokoro-82M	9	54	General use
hexgrad/Kokoro-82M-v1.1-zh	3	103	Chinese optimized

Voice Examples

Language	Female	Male
🇺🇸 American English	af_heart, af_bella, af_nicole	am_michael, am_fenrir
🇬🇧 British English	bf_emma, bf_isabella	bm_george, bm_fable
🇨🇳 Chinese	zf_xiaobei, zf_xiaoyi	zm_yunjian, zm_yunyang
🇯🇵 Japanese	jf_alpha, jf_tebukuro	jm_kumo
🇪🇸 Spanish	ef_dora	em_alex
🇫🇷 French	ff_siwis	-

📚 API Documentation

• Swagger UI: http://localhost:8300/docs
• ReDoc: http://localhost:8300/redoc

🏗️ Project Structure

kokoro/
├── docker/
│   ├── server.py        # FastAPI server
│   ├── ui_template.py   # Web UI
│   └── mcp_server.py    # MCP tools
├── kokoro/              # Core TTS library
├── Dockerfile
├── docker-compose.yml
└── README.md

🛠️ Tech Stack

• Backend: FastAPI, Uvicorn
• TTS Engine: Kokoro-82M (StyleTTS 2)
• Deep Learning: PyTorch, CUDA
• Container: Docker, NVIDIA Container Toolkit

🤝 Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

1. Fork the repository
2. Create your feature branch (git checkout -b feature/amazing)
3. Commit your changes (git commit -m 'Add amazing feature')
4. Push to the branch (git push origin feature/amazing)
5. Open a Pull Request

📖 Documentation

• Streaming Optimization Guide - How we reduced latency from 2s to 50ms

📝 Changelog

v1.1.1 (2025-01) - 🔒 Keep Model Loaded

🆕 New Features

• Added KEEP_MODEL_LOADED environment variable
• When set to true, model stays in GPU memory permanently
• Eliminates cold start delay completely for consistent 51ms TTFB

📊 Latest Performance (2025-01-29)

• Local TTFB: 51-53ms (stable)
• Cloudflare TTFB: 138-178ms
• First chunk: 54ms, 71.5KB

v1.1.0 (2025-01) - 🚀 Streaming Latency Optimization

⚡ Major Performance Improvements

• 40x faster first play - Reduced from 2s+ to ~50ms
• 6x smaller first chunk - Reduced from 436KB to 71.5KB
• 10x faster TTFB - Reduced from ~500ms to ~50ms (local)

🔧 Backend Optimizations

• Split audio by sentence/clause ([.!?。！？,，;；:：]+) instead of newline
• Model warmup on startup - eliminates cold start delay (93ms → 56ms)
• Added X-Accel-Buffering: no and Cache-Control: no-cache headers
• Streaming chunks now generated per sentence for immediate delivery

🎨 Frontend Optimizations

• Non-blocking audio decoding with .then() instead of await
• AudioContext auto-resume for browser autoplay policy
• Immediate playback when first chunk decoded
• Parallel chunk receiving and audio decoding

📊 Performance Metrics Panel (Stream tab)

• Time to First Byte (TTFB) - measures server response time
• Time to First Play - measures actual audio start time
• Total Time - real-time elapsed time counter
• Data Size - total bytes received

🎛️ UI Enhancements

• Added Model selector to Stream, WebSocket, Batch tabs
• Added Voice selector to Stream, WebSocket, Batch tabs
• Added Speed slider to Stream, WebSocket, Batch tabs
• Real-time metrics update during streaming
• Improved status indicators and toast notifications

🐛 Bug Fixes

• Fixed WebSocket sendWS() using wrong model selector
• Fixed Batch tab missing audio playback controls
• Fixed version number display in UI footer

v1.0.0 (2025-01) - 🎉 Initial Release

✨ Core Features

• Beautiful Web UI with 4 tabs (Single, Stream, WebSocket, Batch)
• Full-featured REST API with Swagger/ReDoc documentation
• WebSocket real-time bidirectional TTS
• Streaming audio delivery as chunks generate
• Batch processing for multiple texts

🤖 AI Integration

• MCP Server for AI agent integration (Claude, Cursor, etc.)
• Tool-based TTS generation for AI workflows

🌍 Multi-language Support

• 9 languages: English, Chinese, Japanese, Spanish, French, Hindi, Italian, Portuguese, Korean
• 54+ voices with male and female options
• Multi-model support: Kokoro-82M (general) and Kokoro-82M-v1.1-zh (Chinese optimized)

🚀 Infrastructure

• GPU accelerated with CUDA support
• Automatic GPU memory management with configurable idle timeout
• CPU fallback when GPU unavailable
• Docker containerized deployment

v1.0.0 (2025-01) - 🎉 Initial Release

✨ Core Features

• Beautiful Web UI with 4 tabs (Single, Stream, WebSocket, Batch)
• Full-featured REST API with Swagger/ReDoc documentation
• WebSocket real-time bidirectional TTS
• Streaming audio delivery as chunks generate
• Batch processing for multiple texts

🤖 AI Integration

• MCP Server for AI agent integration (Claude, Cursor, etc.)
• Tool-based TTS generation for AI workflows

🌍 Multi-language Support

• 9 languages: English, Chinese, Japanese, Spanish, French, Hindi, Italian, Portuguese, Korean
• 54+ voices with male and female options
• Multi-model support: Kokoro-82M (general) and Kokoro-82M-v1.1-zh (Chinese optimized)

🚀 Infrastructure

• GPU accelerated with CUDA support
• Automatic GPU memory management with configurable idle timeout
• CPU fallback when GPU unavailable
• Docker containerized deployment

📄 License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.

🙏 Acknowledgements

• hexgrad/Kokoro-82M - The amazing TTS model
• StyleTTS 2 - Model architecture

---

⭐ Star History

Star History Chart

📱 Follow Us