despite being trained to 100k, this is not good sounding modl. All dataset lines were shouting lines, so trying to use calm sounding refrence will make model create strange out-of-character and out of tune sounding  outputs.
the model is still usale, just keep that one thing in mind when working with it.