WFST-BASED STRUCTURAL CLASSIFICATION INTEGRATING DNN ACOUSTIC FEATURES AND RNN LANGUAGE FEATURES FOR SPEECH RECOGNITION