search

openvpi / DiffSinger

An advanced singing voice synthesis system with high fidelity, expressiveness, controllability and flexibility based on DiffSinger: Singing Voice Synthesis via Shallow Diffusion Mechanism
Apache License 2.0
2.62k stars 275 forks source link

i wanna use "forced automatic pitch modem" without F0 truth #92

Closed lvZic closed 1 year ago

lvZic commented 1 year ago

however, following crashed

https://github.com/openvpi/DiffSinger/blob/d9f66c7961a4ee82049c24a99499a4f1966fc1e7/inference/ds_cascade.py#L267

File "openvpi/DiffSinger/modules/naive_frontend/encoder.py", line 72, in forward delta_l = nframes - f0.size(1) AttributeError: 'NoneType' object has no attribute 'size'

seems it stiil use input f0 as model param.

yqzhishen commented 1 year ago

This mode is not suggested because in the current default branch, most models cannot predict f0 themselves, and models which have f0 predictors outputs poor f0. I don't think you will really need to run a model without f0 truth.

F0 predictors are totally removed in refactor-v2 branch, and are re-implemented with more advanced architecture and algorithm in variance branch.

lvZic commented 1 year ago

This mode is not suggested because in the current default branch, most models cannot predict f0 themselves, and models which have f0 predictors outputs poor f0. I don't think you will really need to run a model without f0 truth.

F0 predictors are totally removed in refactor-v2 branch, and are re-implemented with more advanced architecture and algorithm in variance branch.

So i should use variance branch to inference without F0 truth? In fact, i hane tried the author's official version, and the preformance seems not sota.

lvZic commented 1 year ago

This mode is not suggested because in the current default branch, most models cannot predict f0 themselves, and models which have f0 predictors outputs poor f0. I don't think you will really need to run a model without f0 truth.

F0 predictors are totally removed in refactor-v2 branch, and are re-implemented with more advanced architecture and algorithm in variance branch.

I wanna know how the ds_variance.py file get the final wav file by returned dur_pred, f0_pred, variance_pred ?

yqzhishen commented 1 year ago

Non-default branches are not suitable for any user who does not have much coding skills and knowledge about this project. Please wait for the formal release in June.