Kokoro-82M v1.0: architecture based on StyleTTS 2 and ISTFTNet, trained by @rzvzn, published by hexgrad. MLX 8-bit conversion: mlx-community / prince-canuma. Model and voice weights are supplied under Apache-2.0. The pinned MLX revision is 7e173a214392b1e0cb397e71c9745cbd14c06063. Heart (af_heart) and Alpha (hf_alpha) are the selected voices.
The upstream model card acknowledges CC BY training audio: Koniwa tnc, CC BY 3.0 (https://github.com/koniwa/koniwa), and SIWIS, CC BY 4.0 (https://datashare.ed.ac.uk/handle/10283/2353). Retained model cards include the upstream attribution links. No training dataset is distributed here.