A nonuniform sampling approach to isolated word recognition

Abstract

Speech signals are known to consist of both stationary and transient parts. The transient part is invariably shorter than it stationary counterpart. The standard uniform sampling approach for creating word reference patterns for speech recognition will naturally emphasize the longer steady‐state portions more than the short transitional portions. Since the transient portions of words contain valuable consonantal information for recognition, a nonuniform sampling approach seems to be a natural choice for potentially improving the recognizer performance as well as for reducing the overall sampling rate and computational complexity of the recognizer. However, there are some inherent difficulties in this approach. They are (1) artifacts and distortions such as breath noise, mouth clicks, etc. contribute non‐negligible spectral variations and result in over‐sampling in these regions; (2) the inherent variance of the spectral estimate of the transitional regions is much larger than than of a steady‐state region. Nonuniform sampling therefore suffers high uncertainty in its spectral estimate since the high variance regions are sampled more often. Recognition results are presented on a 39‐word, alpha‐digit vocabulary and compared with those from a uniform sampling approach.