Comment by mikemzz1

interesting approach. But training ctc head on top of gemma's audio encoder will most likely produce better results and faster inference