Was there any intent behind the audio encoder representation or was it frozen backbone -> co adaptation?