An Empirical Study of Language Model Integration for Transducer based Speech Recognition