CHANNEL NORMALIZATION TECHNIQUES FOR AUTOMATIC SPEECH RECOGNITION OVER THE TELEPHONE

In this paper we aim to identify the underlying causes that cart expla in the performance of different channel normalization techniques. To t his aim we compared four different channel normalization techniques wi thin the context of connected digit recognition over telephone lines: cepstrum m-an subtraction, the dynamic cepstrum representation, RASTA filtering and phase-corrected RASTA. We used context-dependent and con text-independent hidden Markov models that were trained using a wide r ange of different model complexities. The results of our recognition e xperiments indicate that each channel normalization technique should p reserve the modulation frequencies in the range between 2 and 16 Hz in the spectrum of the speech signals. At the same time, DC components i n the modulation spectrum should be effectively removed. With context- independent models the channel normalization filter should have a flat phase response. Finally, for our connected digit recognition task it appeared that cepstrum mean subtraction and phase-corrected RASTA perf ormed equally well for context-dependent and context-independent model s when equal amounts of model parameters were used. (C) 1998 Elsevier Science B.V. All rights reserved.