How to choose upsampling rte for conditioning? about waveglow

[npuichigo]

@ajinkyakulkarni14 - If the frame rate of your local condition is 1000hz, then the upsampling factor should be 16khz / 1khz = 16. For another example, if you use 5ms frames as local condition, then the upsampling factor would be 16khz * 5ms = 80. I use 200 for 12.5ms mel spectrograms by default, so I think your should adjust the transposed convolution layers to meet your requirements.

[ajinkyakulkarni14]

Thank you for the reply :)

[ajinkyakulkarni14]

@npuichigo ! Also, I would like to know how you set the squeeze factor on local conditioning ?

Also, I have question, if I add nn.Embedding in Upsampling model, do I have to calculate the log Det of it ?

Regards
Ajinkya

[npuichigo]

@ajinkyakulkarni14 After upsampling of local condition, the frame rate of your audio and your local condition should be the same. So I think it's suitable to use the same squeeze factor of the audio samples, or maybe you can reduce both the upsampling factor and the squeeze factor.

Local condition doesn't affect the form of probability transform, so you can add embeddings without having to calculate the log det of it.

[ajinkyakulkarni14]

@npuichigo Thanks alot for the answer !

Ajinkya Kulkarni