2.0 Pre release

Diffusion SVC v2.0 is coming soon.

This model is a combination of NaiveV2, NaiveV2Diff, and Vocoder.
NaiveV2 and NaiveV2Diff is a cascaded training LYNXNet front stage and LYNXNet diffusion model.
They are extremely small in size and highly efficient.

You can train such a model by using configs/config_naivev2diff_comb.yaml and combining them with a fine-tuning vocoder using combo.py.

fine-tuning vocoder ：https://github.com/openvpi/SingingVocoders

1.0 Demo Combo Model

Shallow diffusion model:
k_step_max=100
unit encoder: contentvec768l12
training 600000 steps without pretrain model
network: 512*20
speaker1: opencpop
speaker2: kiritan
Naive model:
unit encoder: contentvec768l12
training 200000 steps without pretrain model
speaker1: opencpop
speaker2: kiritan