BEEP sound with Multiband-Melgan

[tekinek]

I am having a trouble with multiband-melgan:

1. As shown in the following figures, there is a consistant "BEEP" sound in the background of the generated speech. Keep training couldn't get rid of it. The dataset is clean and single-speaker.

2. I have trained the multiband-melgan in an another repo with the same dataset and audio settings (and same model params to my eyes), whose combination with Coqui's tacotron-DDC results in much better sound quality with very silent backgroud:
   https://github.com/TensorSpeech/TensorFlowTTS/blob/34358d82a4c91fd70344872f8ea8a405ea84aedb/examples/multiband_melgan/conf/multiband_melgan.v1.yaml
   But it is a tensorflow implentation that comsumes ~5G GPU memory at inference while Coqui's mb-melgan has much smaller memory footprint and fast inference on GPU.
   I hope someone help me to find out what is different in two implementation (or in settings), so that I may tweak Coqui's mb-melgan work as good as tensorflow-tts. In both cases I have used default configs of ljspeech.

Thanks!

[Amith5970]

have you resolved it?