lucasjinreal/Kokoros

What it solves

Kokoros is a high-performance Rust implementation of the Kokoro TTS model, designed to provide extremely fast text-to-speech synthesis with high-quality audio output. It eliminates the need for heavy Python dependencies by providing a standalone binary and a Rust crate for easy integration into other applications.

How it works

The project uses the Kokoro 82M parameter model via ONNX Runtime to perform inference. It includes a built-in phonemizer and tokenizer (supporting Espeak-ng) to handle text processing end-to-end. The system can be used as a command-line tool (koko), a Rust library, or an OpenAI-compatible HTTP API server that supports streaming audio generation and parallel processing to optimize throughput.

Who it’s for

• Developers wanting to embed high-quality TTS in Rust applications.
• Users seeking a fast, lightweight CLI tool for audio synthesis.
• Engineers building AI agents or digital humans that require low-latency, streaming audio responses.

Highlights

• Insanely Fast Inference: Optimized for speed in Rust, with support for CUDA and WASM.
• OpenAI Compatible: Provides an API server that mimics the OpenAI TTS endpoint for easy integration.
• Streaming Support: Supports both CLI and HTTP API streaming for immediate audio playback (low time-to-first-audio).
• Style Mixing: Allows blending different voice styles (e.g., af_sky.4 + af_nicole.5) for custom voice effects.
• Word-level Timestamps: Can generate .tsv sidecar files providing precise start and end times for each word.
• End-to-End: Includes internal phonemization, removing the need for external dependencies.