Warning: ./VCTK-Corpus/wav48/p283/p283_129.wav was ignored as it contains only silence.

[muthiyanbhushan]

Hello,

I am trying to run the wavenet network according to the instructions. I get below warning for all the .wave files and after pre-processing I get the error. I have downloaded VCTK-Corpus(10.4 GB) and placed it in tensorflow-wavenet/ directory.

Warning: ./VCTK-Corpus/wav48/p283/p283_129.wav was ignored as it contains only silence. Consider decreasing trim_silence threshold, or adjust volume of the audio.

Can someone tell what does this warning means ?

Thank you.

[skywolf829]

I think there is some processing to determine is a .wav file is empty. Since a large portion of some files are static white noise, the algorithm might be thinking they're empty and saying this.

When running the training file, try --noise_threshold=0.1 or lower, and you shouldn't see this as often. This will, however, sometimes cause the network to be trained on white noise, which is far from the human voice you want it to understand.

I've been having an issue with this as well, and I think that's why my loss is staying >2 after 30k iterations. I fear my nosce threshold is too low. Perhaps ibab's 0.3 as the default works best even though it excludes files often.

[krupach]

I'm having the same problem but it's very surprising that this issue got raised only a few weeks ago for the first time.

[aegyoh]

@skywolf829 That seems to have lowered the amount of warning, thanks. It is silence_threshold though and not noise_threshold.

[jawhster]

This issue was discussed last year also https://github.com/ibab/tensorflow-wavenet/issues/270