investigation of acoustic modeling techniques for lvcsr…：声学建模技术为系统研究….pdf

INVESTIGATION OF ACOUSTIC MODELING TECHNIQUES FORLVCSR SYSTEMS X. Liu, M. J. F. Gales, K. C. Sim &K. Yu Cambridge University Engineering Dept, Trumpington St., Cambridge, CB2 1PZ . Email: { xl207,mjfg,kcs23,ky219 } ***@ ABSTRACT This paper describes the use of several advanced acoustic mod- eling techniques for the 2004 CU-HTK large vocabulary speech recognition systems. These techniques include Gaussianization for speaker normalization, discriminative Cluster Adaptive Training (CAT), Subspace for Precision And Mean (SPAM) modeling of in- verse covariances, and plexity control. Acous- tic models featuring these techniques were integrated into a state- of-the-art 10 real-time multi-pass system with sophisticated adap- tation for performance evaluation. Experimental results are pre- sented on both broadcast news (BN) and conversational telephone speech (CTS) transcription tasks. 1. INTRODUCTION For many years automatic transcription of broadcast news (BN) and conversational telephone speech (CTS) data have been the two main tasks for the munity of large vocabulary contin- uous speech recognition (LVCSR). Due to the dif ?culty of these tasks, a variety of modeling techniques have been developed to make these systems able to model plex data and more robust to changes in acoustic environment. In this paper several advanced modeling techniques are investigated in the fra