Towards Effective and Compact Contextual Representation for Conformer Transducer Speech Recognition Systems