Modelpublic

wabibito/Onyx-Kokoro-82M

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

1likes43downloads

Model Card

Onyx-Kokoro-82M

A mixed-precision MLX quantization of Kokoro-82M (StyleTTS2-derived, Apache-2.0), packaged for [Onyx](https://github.com/) — an on-device iOS AI assistant. Used by OnyxTTS, Onyx's own from-scratch Swift + MLX text-to-speech engine, to speak assistant replies in a natural neural voice fully on-device (Apple's AVSpeechSynthesizer is the fallback).

What this is

—Mixed 4-bit quant (group size 64): the PL-BERT (ALBERT) text encoder's Linear weights are 4-bit; the entire vocoder path — the prosody predictor (LSTMs + AdaIN convs → pitch/energy) and the iSTFTNet generator — stays fp16. The vocoder directly shapes the waveform and is quant-sensitive, so it is deliberately NOT quantized. ~330 MB fp32 → 148 MB (model.safetensors).
—Voice packs (voices/*.safetensors): 54 voices (af/am American, bf/bm British, plus jf/jm/zf/zm/ ef/em/ff/hf/hm/if/im/pf/pm for other languages). Each is a [N, 1, 256] style tensor indexed by phoneme-sequence length.
—config.json carries the standard Kokoro config plus a quantization block ({group_size: 64, bits: 4}) so the loader dequantizes the right layers.

Provenance

Weights derived from prince-canuma/Kokoro-82M (the mlx-audio mirror of hexgrad/Kokoro-82M) via Onyx's quantize_kokoro.py. Verified to load + synthesize valid 24 kHz audio end-to-end through OnyxTTS on Apple Silicon. Apache-2.0, same as the base model.