4-1-train-vits2.md

Part 4.1: e.g. VITS 2 training

GIVE UP because training too long, switch to Matcha-TTS much faster

text-to-speech using https://github.com/p0p4k/vits2_pytorch

edit its vits2_pytorch/requirements.txt to include torch with cuda, underthesea, num2words
(require MSVC to build Monotonic Alignment Search)

replace content of file vits2_pytorch/text/cleaners.py with my file speech-synthesis-ngngngan/scripts/_cleaner.py

edit file vits2_pytorch/utils.py:

• change logging.DEBUG to logging.WARNING
• in function remove_old_checkpoints remove print because it disturbs tqdm progress bar

because NCCL not support on windows, must edit file vits2_pytorch/train.py:

• replace mp.spawn(…) with run(0, n_gpus, hps)
• remove dist.init_process_group(…)
• remove any DDP()
• replace ███.module.███ with ███.███

train: python train.py --model vits2_ngngngan --config ../speech-synthesis-ngngngan/data/vits2_ngngngan_nosdp.json

config file vits2_ngngngan_nosdp.json inspired from

• https://github.com/p0p4k/vits2_pytorch/blob/main/configs/vits2_ljs_nosdp.json
• use newer implementation of duration discriminator (dur_disc_2) in p0p4k/vits2_pytorch#59

continue training from the only vietnamese checkpoints i found, shared in p0p4k/vits2_pytorch#10 (comment) (trained 111k steps but didn’t have duration discriminator) ⇒ 1 h/epoch instead of 2 h/epoch if train from scratch

logs and checkpoints saved in folder vits2_pytorch/logs/vits2_ngngngan