config.json
| 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 | "tilt_norm": { |
| 19 | "mean_db_per_oct": -7.8, |
| 20 | "sd": 1.8 |
| 21 | }, |
| 22 | "blend": { |
| 23 | "packs": 28, |
| 24 | "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", |
| 25 | "gate_st": 4.0, |
| 26 | "stat_scale": [1.0, 1.1, 0.8], |
| 27 | "grade_pen": 1.0, |
| 28 | "grades": "overall_grade per pack from Kokoro-FastAPI voice_grades.json, A = 4.0, C = 2.0; weight on packs below C is penalized" |
| 29 | }, |
| 30 | "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", |
| 31 | "prosody_head": { |
| 32 | "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", |
| 33 | "output": "delta added to every row of the predictor half after the blend", |
| 34 | "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" |
| 35 | } |
| 36 | } |
| 37 | |