Kokoro v1.0 — ONNX + Voice Bundle
Kokoro-82M is an 82M-parameter text-to-speech model that punches well above its weight class. This repo bundles the ONNX checkpoint with the 11 official voice embeddings so users can download once and have the full vocal range ready.
Not converted locally — Kokoro publishes ONNX as its primary distribution format. This is a curated re-host with the voices co-located for convenience.
Credit: hexgrad (Kokoro).
What this repo contains
kokoro-v1.0.onnx # 311 MB — the TTS model
kokoro-voices/
af.bin # default American Female (alias for af_bella)
af_bella.bin
af_nicole.bin
af_sarah.bin
af_sky.bin
am_adam.bin # American Male
am_michael.bin
bf_emma.bin # British Female
bf_isabella.bin
bm_george.bin # British Male
bm_lewis.bin
Total: ~320 MB.
Voice naming convention: {a,b}{f,m}_<name>.bin where a = American, b = British, f = female, m = male.
How to use
Kokoro takes a tokenized phoneme sequence as input. You'll typically pair it with a phonemizer library (the upstream Python uses phonemizer + espeak-ng).
import onnxruntime as ort
import numpy as np
sess = ort.InferenceSession("kokoro-v1.0.onnx")
voice = np.fromfile("kokoro-voices/af_bella.bin", dtype=np.float32).reshape(-1, 1, 256)
# tokens: int64 array of phoneme IDs from your phonemizer
# speed: float in [0.5, 2.0], 1.0 = natural
audio = sess.run(None, {
"tokens": tokens,
"style": voice[len(tokens)], # voice embedding indexed by token length
"speed": np.array([1.0], dtype=np.float32),
})[0] # → float32 waveform at 24 kHz
See hexgrad/Kokoro-82M for the canonical reference implementation.
License
Apache-2.0 — same as upstream. LICENSE file included.