kokoro-rust

Rust port of kokoro-tts — a fast wrapper around the Kokoro-82M ONNX text-to-speech model.

~1.7s to generate ~9.0s of audio (ONNX Runtime on M1 Mac)

Prerequisites

• cargo (Rust toolchain)
• espeak-ng (phonemization via FFI)
• uv (voice pack conversion script)
• curl (model/voice download)

Install

git clone https://github.com/RohanAwhad/kokoro-rust.git
cd kokoro-rust

Setup

Download the ONNX model:

curl -L -o ~/Library/Caches/kokoro-tts/kokoro-v0_19.onnx \
  "https://huggingface.co/hexgrad/Kokoro-82M/resolve/main/kokoro-v0_19.onnx"

Linux: cache dir is ~/.cache/kokoro-tts/ instead of ~/Library/Caches/kokoro-tts/.

Download and convert a voice pack:

curl -L -o /tmp/af_sky.pt \
  "https://huggingface.co/hexgrad/Kokoro-82M/resolve/main/voices/af_sky.pt"

mkdir -p ~/Library/Caches/kokoro-tts/voices/
./scripts/convert_voices.py /tmp/af_sky.pt \
  ~/Library/Caches/kokoro-tts/voices/af_sky.kokoro

Usage

CLI (streams audio in real-time):

cargo run --bin kokoro-tts -- "Hello world!"

Write to WAV:

cargo run --bin kokoro-tts -- -o output.wav "Hello world!"

Python API:

from kokoro import Kokoro  # Rust via PyO3 (future)

Rust API:

use kokoro::{Kokoro, Voice};

let kk = Kokoro::new();
let audio = kk.generate("Hello world!", Voice::AfSky)?;
// audio.sample_rate == 24000
// audio.data: Vec<f32>

Voices

Voice	Language
af	American female
af_bella	American female
af_sarah	American female
am_adam	American male
am_michael	American male
bf_emma	British female
bf_isabella	British female
bm_george	British male
bm_lewis	British male
af_nicole	American female
af_sky	American female (CLI default)

Build

cargo build --release
# Binary: target/release/kokoro-tts