config.json
1.7 KB · 38 lines · json Raw
1 {
2 "name": "inno_ref01",
3 "version": "0.2.0",
4 "model_type": "kokoro-inno-reference-tuner",
5 "base_model": "hexgrad/Kokoro-82M",
6 "base_model_file": "kokoro-v1_0.pth",
7 "kokoro_version": "0.9.4",
8 "speaker_encoder": "baked enc.* in model.safetensors: wespeaker-resnet34 distilled from microsoft/unispeech-sat-base-plus-sv",
9 "embedding_dim": 512,
10 "style_dim": 256,
11 "voice_pack_shape": [
12 510,
13 1,
14 256
15 ],
16 "sample_rate": 24000,
17 "weights": "model.safetensors",
18 "onnx": "model.onnx: speaker encoder + style head, input fbank [1, frames, 80] float32 (inno_kokoro.fbank), outputs embedding [1, 512], style [1, 256]",
19 "tilt_norm": {
20 "mean_db_per_oct": -7.8,
21 "sd": 1.8
22 },
23 "blend": {
24 "packs": 28,
25 "stats": "F0 mean (st), F0 spread (st), syllables per second, measured once on each stock pack's render of a fixed passage; baked as blend.stats",
26 "gate_st": 4.0,
27 "stat_scale": [1.0, 1.1, 0.8],
28 "grade_pen": 1.0,
29 "grades": "overall_grade per pack from Kokoro-FastAPI voice_grades.json, A = 4.0, C = 2.0; weight on packs below C is penalized"
30 },
31 "style_head": "heads.style.* / heads.tilt.weight: speaker embedding -> timbre half, trained on LibriTTS-R + VoxPopuli-en + Emilia-YODAS EN speakers with the predictor half fixed to the blend",
32 "prosody_head": {
33 "inputs": "reference F0 mean and sd (st), Praat over 6 s chunks; baked as head.W [3, 128], standardized by head.mu / head.sd, last row = bias",
34 "output": "delta added to every row of the predictor half after the blend",
35 "fit": "ridge on 426 speakers (LibriTTS-R, VoxPopuli-en, Emilia-YODAS EN) whose deltas were optimized through the frozen predictor to match F0 mean, spread, IQR, pause fraction and speech time"
36 }
37 }
38