Another issue of model failing to learn a sine wave.

In https://github.com/ibab/tensorflow-wavenet/issues/141, the main discussion is about the slightly change in wave form when generating sine wave using quantization 256, which according to me is not the cause of noise. The trimming method is also not a problem because I disabled it in my test.

I tried to make the model learn to create the sin wave, but it turns out that it generates a totally noise audio. Here is a picture which I got from the plot generated_audio[7900:8100]. I use seed of the same sin wave audio file to eliminate the effect of initial seed. I even let the optimizer run through 500 steps to get the loss=0.0001, so the model overfits. As you can see from the picture, the first 100 points is the sine wave (a slightly change due to quantization) getting from the seed, but after that, when start to generate, the model fails.

Edit: I forget to test the fast_generation method. It can confirm that it is the cause of the problem. Relevant issue is from here: https://github.com/ibab/tensorflow-wavenet/issues/125.

[jyegerlehner]

@fehiepsi mu-law encoding is very low resolution at the extreme ends of the range. If you scale it from -0.5 to 0.5 instead of -1.0 to 1.0 it will look (and sound) better.

[fehiepsi]

@Zeta36 : it is not perfect, here is a picture