kokoroxide [WIP]

A high-performance Rust implementation of Kokoro TTS (Text-to-Speech) synthesis, leveraging ONNX Runtime for efficient neural speech generation. Uses espeak-ng for text-to-phoneme conversion, with built-in conversion logic into Misaki phoneme notation expected by Kokoro models. Distributed under a dual MIT/Apache-2.0 license to match the broader Rust ecosystem.

Note: Currently only supports and has been tested with American English. Contributions for different languages are very welcome!

Features

• 🎨 Voice Style Control - Customize voice characteristics with style vectors
• 🔤 Phoneme Support - Direct phoneme input for precise pronunciation control
• ⚡ Speed Control - Adjust speech rate dynamically
• 🔧 Flexible API - Multiple generation methods for different use cases

Installation

Add this to your Cargo.toml:

[dependencies]
kokoroxide = "0.1.1"

Quick Start

use kokoroxide::{load_voice_style, KokoroTTS, TTSConfig};

fn main() -> Result<(), Box<dyn std::error::Error>> {
    // Configure the ONNX model + tokenizer that Kokoro requires.
    // These files live outside the crate; download them from Kokoro's distribution (https://huggingface.co/onnx-community/Kokoro-82M-v1.0-ONNX).
    let config = TTSConfig::new("path/to/kokoro.onnx", "path/to/tokenizer.json")
        .with_sample_rate(24000)
        .with_max_tokens_length(512)
        .with_graph_optimization_level(kokoroxide::GraphOptimizationLevel::Disable);

    // Build the speech engine with the explicit configuration so advanced knobs are available.
    let tts_service = KokoroTTS::with_config(config)?;

    // Load a voice style vector (.bin) that controls prosody and speaker identity.
    let voice = load_voice_style("path/to/voice.bin")?;

    // Generate speech at 1.0x speed for the requested text.
    let text = "Hello, this is a text-to-speech synthesis example.";
    let audio = tts_service.generate_speech(text, &voice, 1.0)?;

    // Persist the synthesized waveform to a WAV file for playback.
    audio.save_to_wav("path/to/output.wav")?;

    Ok(())
}

For a complete runnable example pointing at real assets, see the kokoroxide-demo sample project in this workspace (kokoroxide-demo/src/main.rs).

API Overview

Core Types

KokoroTTS

The main TTS engine that handles text-to-speech conversion.

// Create with default config
let tts = KokoroTTS::new(model_path, tokenizer_path)?;

// Create with custom config
let config = TTSConfig::new(model_path, tokenizer_path)
    .with_max_tokens_length(128)
    .with_sample_rate(24000);
let tts = KokoroTTS::with_config(config)?;

VoiceStyle

Represents voice characteristics as a style vector. Voice files contain multiple style vectors indexed by token length.

// Load from binary file
let voice = load_voice_style("voice.bin")?;

// Create custom voice with vector size
let custom_voice = VoiceStyle::new(vec![0.1, 0.2, ...], 256);

GeneratedAudio

Contains the generated audio samples and metadata.

let audio = tts.speak("Hello!", &voice)?;
println!("Duration: {} seconds", audio.duration_seconds);
println!("Sample rate: {} Hz", audio.sample_rate);
audio.save_to_wav("output.wav")?;

Generation Methods

1. Simple Text-to-Speech

let audio = tts.speak("Hello, world!", &voice)?;

2. With Speed Control

let audio = tts.generate_speech("Speak faster!", &voice, 1.5)?; // 1.5x speed

3. From Phonemes

let audio = tts.generate_speech_from_phonemes("həˈloʊ wɜːld", &voice, 1.0)?;

4. From Token IDs

let tokens = vec![101, 2234, 1567, 102]; // Pre-tokenized input
let audio = tts.generate_from_tokens(&tokens, &voice, 1.0)?;

Configuration

TTSConfig Options

use ort::GraphOptimizationLevel;

let config = TTSConfig::new(model_path, tokenizer_path)
    .with_max_tokens_length(512)    // Maximum token sequence length
    .with_sample_rate(24000)        // Audio sample rate in Hz
    .with_graph_optimization_level(GraphOptimizationLevel::Level3); // ONNX graph optimization

Graph Optimization Levels

The with_graph_optimization_level() method allows you to control ONNX Runtime's graph optimization:

• GraphOptimizationLevel::Disable - No optimizations
• GraphOptimizationLevel::Level1 - Basic optimizations
• GraphOptimizationLevel::Level2 - Extended optimizations
• GraphOptimizationLevel::Level3 - Maximum optimizations (default)

System Requirements

Prerequisites

1. Rust 1.70+

2. espeak-ng (required for text-to-phoneme conversion):

   • Ubuntu/Debian: sudo apt-get install espeak-ng libespeak-ng-dev
   • macOS: brew install espeak-ng
   • Windows: Download from espeak-ng releases
   • Arch Linux: sudo pacman -S espeak-ng

3. ONNX Runtime (automatically downloaded via ort crate)

4. Kokoro model files:

   • Model file (e.g., kokoro-v0_19.onnx)
   • Tokenizer configuration (tokenizer.json)
   • Voice style files (.bin format)
   • Downloaded at runtime or managed outside the crate package to keep the published crate lightweight

Build Configuration

The crate automatically links to espeak-ng based on your platform:

• macOS: Looks for espeak-ng in /opt/homebrew/lib (Homebrew default)
• Linux: Uses system library paths

If espeak-ng is installed in a non-standard location, you may need to set:

export LD_LIBRARY_PATH=/path/to/espeak-ng/lib:$LD_LIBRARY_PATH  # Linux
export DYLD_LIBRARY_PATH=/path/to/espeak-ng/lib:$DYLD_LIBRARY_PATH  # macOS

Environment Variables

• DEBUG_PHONEMES - Enable phoneme debugging output:

  DEBUG_PHONEMES=1 cargo run
  

  This will print:

  • Input text
  • Espeak IPA output
  • Converted Misaki phonemes

• DEBUG_TOKENS - Enable token debugging output:

  DEBUG_TOKENS=1 cargo run
  

  This will print:

  • Generated token IDs array

• DEBUG_TIMING - Enable performance timing logs:

  DEBUG_TIMING=1 cargo run
  

  This will print:

  • Phoneme tokenization time
  • Espeak IPA conversion time
  • Total tokenization time

• All debug modes:

  DEBUG_PHONEMES=1 DEBUG_TOKENS=1 DEBUG_TIMING=1 cargo run
  

Model Files

Download the Kokoro model files from the official repository:

• Model: Kokoro-82M ONNX
• Required files:
  • *.onnx - The model file
  • tokenizer.json - Tokenizer configuration
  • Voice files (*.bin) - Style vectors for different voices
  • Provide these assets at runtime (they are not packaged with the crate to keep the published tarball lightweight)

Examples

Basic TTS Application

use kokoroxide::{KokoroTTS, load_voice_style};

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let tts = KokoroTTS::new("model.onnx", "tokenizer.json")?;
    let voice = load_voice_style("voice.bin")?;

    let text = "Welcome to kokoroxide TTS!";
    let audio = tts.generate_speech(text, &voice, 1.0)?;
    audio.save_to_wav("welcome.wav")?;

    println!("Generated {} seconds of audio", audio.duration_seconds);
    Ok(())
}

License

Licensed under either of:

• Apache License, Version 2.0 (LICENSE-APACHE or http://www.apache.org/licenses/LICENSE-2.0)
• MIT license (LICENSE-MIT or http://opensource.org/licenses/MIT)

at your option.

Contributing

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

Acknowledgments

This project implements the Kokoro TTS model in Rust, providing a high-performance alternative to Python implementations.