Dataset is about ~2000 hours of speech and vocals
Supported languages (english or spanish?) who ever moves first is:
~800 hrs of English (with vast verity of speakers and every emotion)
~200 Spanish
~42 French
~188 Russian
~70 Arabic
~140 Japanese
~70 Chinese (Mandarin)
~80 Korean
~30 Hindi
~53 Indonesian
~30 Tagalog
~40 Portuguese
~35 German
~190 singing (all languages)
common language (I don't remember how much data was there)
Type: big-base for finetuning
Batch: 2-40-80
fp32
Sampling frequency: 32k 40k
Total steps count: 371406
Hardware used:
1 - h100, 4 - L40s
Expected release date - 22 july
()