A Deliberation-based Joint Acoustic and Text Decoder