On the Effects of Varying Analysis Parameters on an LPC-Based Isolated Word Recognizer

When faced with the problem of building hardware for a speech processing system, the practical problems of deciding how to implement the system are often solved based on insufficient information of the effects of system parameters on performance. Generally the hardware designer is given a "working system" and asked to devise hardware that performs the same signal processing operations. The designer 893 often sees potential reductions in hardware complexity (price, etc.), but without a good understanding of the tradeoffs between complexity and performance, he cannot utilize his design knowledge in an efficient manner. The situation described above is applicable to a number of areas of speech processing. This is especially the case for speech recognition, in which performance scores for a number of different systems have been reported, but for which there is no good experimental data showing how performance degrades (or improves) as system variables are changed in value. Perhaps the closest that investigators have come to obtaining such performance data are the studies by White and Neely comparing two feature sets (linear predictive parameters and bandpass filter parameters) and two time warping methods (linear warping and dynamic time warping), and the one by Silverman and Dixon, comparing spectral analysis and classification techniques. In this paper, we present results of a systematic study of the effects on performance of the parameters of a linear predictive coding (LPC)based, isolated word recognition system.