Streaming-KokoroJS

About

Unlimited text-to-speech using Kokoro-JS, 100% local, 100% open source

Online Demo (that does not send any user text data to any server.)

https://rhulha.github.io/StreamingKokoroJS/

Overview

Streaming-Kokoro is a web-based text-to-speech application that leverages the Kokoro-82M model to generate high-quality speech audio entirely in the browser. The application runs completely in the browser without requiring any server-side processing or API calls, ensuring privacy and offline functionality.

Features

• 100% Client-Side Processing: All text-to-speech conversion happens locally in your browser
• WebGPU Acceleration: Automatically uses WebGPU for faster processing when available, with WASM fallback
• Streaming Audio Generation: Processes text in chunks and streams audio as it's generated
• Smart Text Chunking: Intelligently splits text to maintain natural speech patterns
• Multiple Voice Styles: Supports various voice styles for different languages
• Audio Download: Save generated audio to disk
• Fully Open Source: Every component is open source and freely available

Technical Details

• Uses the Kokoro-82M-v1.0-ONNX model (~300MB, cached after first load)
• Employs Web Workers for non-blocking UI during speech generation
• Automatically detects hardware capabilities and selects optimal processing mode:
  • WebGPU acceleration on compatible browsers/devices
  • WebAssembly (WASM) fallback on other devices
• Sample rate: 24kHz for high-quality audio output

Getting Started

1. Clone this repository
2. Serve the files using a local web server
3. Open the application in a modern browser (Chrome/Edge recommended for WebGPU support)
4. Enter or paste text into the text area
5. Click "Play" to stream the audio or "Download" to save it to disk

Local Model

The application can use a local model instead of downloading from HuggingFace. Example:

if (self.location.hostname === "localhost") {
  env.allowLocalModels = true;
  model_id = "./my_model/";
}

Model Information

This project uses the Kokoro-82M-v1.0-ONNX model from Hugging Face, which provides high-quality text-to-speech capabilities in a relatively compact package suitable for browser-based applications.

Browser Compatibility

• Recommended: Chromium-based browsers with WebGPU support (Chrome, Edge, etc.)
• Compatible: Any modern browser with WebAssembly support. (FireFox seems not to work at the moment.)

License

This project is open source under the Apache 2 License.

Acknowledgments

• Kokoro is an open-weight TTS model with 82 million parameters.
• Hugging Face Transformers.js for the browser-based ML framework
• Kokoro-82M Model for the TTS capabilities