problem on generate only noise

[wl3b10s]

with default setting,

simply run train on VCTK without extra global condition by : python train.py

then python generate.py --wav_out_path=testresult3900.wav --samples 320000 logdir/train/2020-10-22T13-35-57/model.ckpt-3900

seems doesn't generate voice wav but pure noise.

is there some more setting or condition needs to be added?

thanks.

[tanzyy96]

I'm also facing the same issue. I kept the model training for somewhere less than a week.

[his0car]

I took like 30~ hours to train 9000 steps and this is what came out: https://vocaroo.com/1kYnhSXQBGaN

This haunts me in my dreams. Jokes aside, the generated audio linked in the README was generated with 80k Steps, so I wouldn't expect 3900 steps to suffice

[SerafimC]

@PascalSchimmler thank you for the information. I was looking for some notion of how many steps is necessary to create good outputs.

[magiccpp]

This is my trained model with 80000 steps, https://drive.google.com/file/d/1MjCVFNw8XS0HgmPD0fthscxLMdGtB2cr/view?usp=share_link The generated wave file: generated_seed.wav.gz

[ZhangXiaoying0116]

@smartestrobotdai
Can you share your trained checkpoint with me? I can't find this link https://drive.google.com/file/d/1MjCVFNw8XS0HgmPD0fthscxLMdGtB2cr/view?usp=share_link