mikemzz1··on An hour of audio transcribed in about 13s on MacBook Airinteresting approach. But training ctc head on top of gemma's audio encoder will most likely produce better results and faster inference