software-mansion/react-native-executorch-kokoro

Introduction

This repository hosts the Kokoro model for the React Native Executorch library. It can perform speech synthesis in 8 different languages, including fine-tuned Polish and German.

The models support input shape dynamism and cover the input range of 1 up to 128 tokens.

Additionally, the repository contains essential resources for G2P (grapheme-to-phoneme) preprocessing required by the Kokoro model, including simple word-by-word phonemization models (also in ExecuTorch format).

If you'd like to run these models in your own ExecuTorch runtime, refer to the official documentation for setup instructions.

Compatibility

These models were exported using v1.0.0 version of ExecuTorch and no forward compatibility is guaranteed. Older versions of the runtime may not work with these files.

The models are intended to be used within the React Native ExecuTorch package. If you want to use them outside the package, make sure your runtime is compatible with the ExecuTorch version used to export the .pte files and follow the example script to run the models.

Repository Structure

.
├── config.json              # Backend-agnostic model manifest
├── phonemizer/              # G2P data files + per-language phonemizer models
├── voices/                  # Pre-computed speaker embeddings
└── xnnpack/                 # XNNPACK-exported .pte files, one directory per model variant
    ├── standard/
    │   ├── config.json
    │   ├── duration_predictor_std_xnnpack_fp32.pte
    │   └── synthesizer_std_xnnpack_fp32.pte
    ├── german/
    │   ├── config.json
    │   ├── duration_predictor_de_xnnpack_fp32.pte
    │   └── synthesizer_de_xnnpack_fp32.pte
    └── polish/
        ├── config.json
        ├── duration_predictor_pl_xnnpack_fp32.pte
        └── synthesizer_pl_xnnpack_fp32.pte

• phonemizer - data files required by the Phonemis package - responsible for input preprocessing part of React Native ExecuTorch Kokoro pipeline.
• voices - a collection of pre-computed speaker embeddings used by the Kokoro model to synthesize speech with specific vocal characteristics.
• xnnpack - exported, XNNPACK-optimized Kokoro runtime modules. Each variant directory ships a config.json manifest describing its .pte files and their method signatures.

Each variant is composed of two sub-models that run sequentially:

1. Duration predictor — predicts per-token durations and the prosody features d. Exported as a bundled program exposing forward_32, forward_64 and forward_128, one method per padded token bucket.
2. Synthesizer — encodes text, predicts F0/N and decodes the waveform. Exported with dynamic shapes behind a single forward method.