They say the models are under 3b parameters. If only for voice generation it sounds pretty good, no ?