Core ML conversions of hexgrad/Kokoro-82M
(f3ff357) made from its PyTorch weights, with the data an app needs to run
it. Apache-2.0; see LICENSE and NOTICE for attribution and the changes made.
| File | Contents |
|---|---|
Kokoro_Predictor_fp16.mlpackage.zip |
predictor: input_ids [1,N] int32 (N up to 256, unpadded), ref_s_style [1,128] -> duration, d_for_align, t_en |
Kokoro_Decoder_fp16.mlpackage.zip |
functions decoder_128, decoder_256, decoder_512: en_aligned, asr_aligned, ref_s [1,256] -> audio (frames x 600 samples, 24 kHz) |
kokoro_vocab.json |
phoneme -> token id |
voice_*.bin |
510 x 256 float32 style table for af_heart, af_bella, am_michael, bf_emma, bm_george |
us_gold.json, us_silver.json |
misaki 0.9.4 American-English dictionaries |
sources.json |
where each data file came from, with SHA-256s |
conversion/ |
the scripts that made the archives |
Float16 weights, float32 compute, intended for CPU. Checked against Kokoro's PyTorch output: rounded durations equal on every token; decoder spectral convergence about 0.02-0.03 (float16 against float32 Core ML), measured on five English sentences.
voice_af_bella.bin 522.2 kBvoice_af_heart.bin 522.2 kBvoice_am_michael.bin 522.2 kBvoice_bf_emma.bin 522.2 kBvoice_bm_george.bin 522.2 kBhexgrad/Kokoro-82M