Yoo, this is insanely good

how many epochs / how long was the audio sample you trained it with?

Sorry for the late reply I didn't get a notification that someone commented on this model. ~ I don't know exactly - the training page doesn't display the filenames correctly they get truncated. Thus I can't verify it with my dataset. Strange enough the training page also lists ALL uploaded dataset files from the model history. But I can say that I nearly filled up the maximum dataset length for premium accounts. Audio in the training dataset is pretty clear and noise free. Its interesting to see that the resulting model is also able to churn out some singing even though it was only trained on speech. ~~~~~~Edit: I must add, that I manually removed all problematic sections in the dataset (music and sound bleed throughs). Not sure if I catched them all. :P