tensorflow-wavenet
tensorflow-wavenet copied to clipboard
ibab
Only noise be generated
Hi, I used the code directly, and I got something just like noise. I executed "python train.py" , and then "python generate.py". Finally I got that file in zip.
Thanks! generated.wav.zip
same problem, looking for help. my generated.wav only have few seconds and noise. this maybe helpful
#218
I get this on my Terminal screen. I am not sure if I am running it properly or I shouldn't see this warning. Warning: onespeaker/Audio/p225/p225_265.wav was ignored as it contains only silence. Consider decreasing trim_silence threshold, or adjust volume of the audio. How would I fix it so I don't get this warning?
How many steps? I takes me about 80000+ before things sound okay. Also use a seed wav file. I seem to get better results.
Change the trim_silence threshold or adjust the audio volume to get rid of the warning.
@devinroth hi, can i have some advise from you, i do the 99999 steps , but could not hear the clear sound.
@devinroth what is seed wav file? I followed the original tutorial from: https://github.com/ibab/tensorflow-wavenet, and the only "factor" I changed is SILENCE_THRESHOLD from 0.3 to 0. The corpus I utilized is from VCTK. :)
@devinroth hey Devin, what loss do you get with 80k+ iterations? Can you post the data, including trim threshold and model config you're using?
Take a snippet of an audio file to jump start the generation. Ideally use something that wasn't used while training.
Devin Roth
On Dec 25, 2017, at 2:46 AM, ChrisChan2013 [email protected] wrote:
@devinroth what is seed wav file? I followed the original tutorial from: https://github.com/ibab/tensorflow-wavenet, and the only "factor" I changed is SILENCE_THRESHOLD from 0.3 to 0. The corpus I utilized is from VCTK. :)
— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.
I believe i got around 3% loss. More than that ended up being over trained.
I'll post some stuff if I get a chance. Probably not for a few weeks though.
Devin Roth
On Dec 25, 2017, at 6:32 AM, Rafael Valle [email protected] wrote:
@devinroth hey Devin, what loss do you get with 80k+ iterations? Can you post the data, including trim threshold and model config you're using?
— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.
What do you mean by 3% loss? Loss should be between 0 and + infinity. Did you train on a single vctk speaker or all speakers?
hi @devinroth I'm also getting the same white noise generated when following your readme.
Sorry but I don't understand what you mean by a seeded wav file? do you mean use a previously generated wav file and put it back into the training as a parameter? Of course that will be me putting a noisey wav file back in.
Could you please give some examples?
Is there any chance that someone can update the readme.md to reflect how to simply use this to generate a working wav file because I followed your readme steps and this is what came out.
So I found the parameter in the generate script. I tried the following:
python generate.py --wav_out_path=generated.wav --wav_seed corpus/VCTK-Corpus/wav48/p252/p252_016.wav --samples 16000 logdir/train/2018-09-14T09-27-10/model.ckpt-138
Now I can almost head a human voice but then load of static after.
The seed is an clip of audio like something you trained with. Not generated. Like a clip of someone talking.
On Sep 14, 2018, at 3:00 AM, KiransHub [email protected] wrote:
hi @devinroth I'm also getting the same white noise generated when following your readme.
Sorry but I don't understand what you mean by a seeded wav file? do you mean use a previously generated wav file and put it back into the training as a parameter? Of course that will be me putting a noisey wav file back in.
Could you please give some examples?
— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.