NERDDISCO's picture
kale-tts-1
f21d6eb
Raw History Blame Contribute Delete
1.71 kB
{
"_license": "SPDX-License-Identifier: Apache-2.0. Copyright 2026 Kale.",
"_about": "Render settings of the release data (render.py, make_*.py). Training data = original Maya-1 audio only, uncut, no post-processing.",
"maya1": {
"repo": "maya-research/maya1",
"revision": "b1b9b38ca9decbfcfd6c98ea93dd502bfd0f9572",
"licence": "Apache-2.0"
},
"snac": {
"repo": "hubertsiuzdak/snac_24khz",
"revision": "d73ad176a12188fcf4f360ba3bf2c2fbbe8f58ec",
"licence": "MIT",
"sample_rate": 24000
},
"engine": {
"vllm": "0.33.0",
"dtype": "bfloat16",
"quant": "",
"gpu": "H100",
"gpu_memory_utilization": 0.8,
"max_model_len": 2400
},
"sampling": {
"temperature": 0.4,
"top_p": 0.9,
"repetition_penalty": 1.1,
"min_tokens": 28,
"max_tokens_cap": 2048,
"max_tokens_base": 220,
"max_tokens_per_word": 60,
"seed": "seed0 + row index (one seed per take)"
},
"render_gpus": 4,
"_render_gpus": "max containers (1 H100 each); the 10-voice build used 1-5",
"per_call": 10000,
"text_pool": {
"source": "Project Gutenberg public-domain books (make_text.py BOOKS), boilerplate stripped, no dialogue/digits/brackets",
"text_mix": {
"1": 0.4,
"2": 0.2,
"3": 0.2,
"4": 0.12,
"5": 0.08
},
"max_words": 62,
"max_words_4_5": 66,
"sentence_words": [
3,
28
],
"seed": 1411,
"ban": "bench lines, held-out test sets and fixed test lines (volcano line, level-guard lines) never appear in training text"
},
"singles_pool": {
"sentence_words": [
3,
28
],
"start": 200000
},
"counts_rule": "women ~9-12k mix + 6-9k singles, men 12-20k mix + 18-30k singles (male takes fail the sentence-end check more often); aim for ~25 h kept per voice"
}