Best Practices for Training Data (corpus)

[taylorbrook]

Is there any advice on how to best prepare the training data?

I am about to do my second traning for the nn~ object and was going to use recordings of myself speaking as the traning data. Should I cut out silences as much as possible? If I train on only speaking rather than singing will this produce poor or chaotic results when attempting a transfer from a sustained voice or instrument?

I have successfully trained a bowed guitar model, but the sounding result wasn't amazing, and my hunch was because of how drone-like on a single pitch the audio used for training data was.

Apologies if this is addressed elsewhere, I made some cursory searches and couldn't find anything.