Constructing new mels as Input about wavernn

[fatchord]

@ZohaibAhmed Yeah, that tacotron2 repo is absolutely fine for creating training mels. I used my own implementation of tacotron for the demo samples.

[G-Wang]

In the preprocessing ipython notebook, you can see the steps used to get Mel spectrlgram from audio wav.

WavRNN is by itself a vocoder, so it only takes in Mel spectrogram. You can pass the Mel spectrogram generated by a text to Mel model, such as tacotron.

You need to make sure that the Mel spectrogram processing is the same for both parts.

[ZohaibAhmed]

@G-Wang Got it. I'll try to use https://github.com/Rayhane-mamah/Tacotron-2 to generate the Mel.

[bjtommychen]

Hi, all
I don't understand what the input is.
for 1/2/3, it's wave data clearly.
but for 4b. it's wave data + melspectrogram. how can feed this two into WaveRNN() of 1/2/3.
Thanks.

[fatchord]

@bjtommychen You're correct, notebooks 1, 2 & 3 are unconditioned. However if shouldn't be too difficult to add the upsampling network from 4b to the original WaveRNN model if you feel so inclined.