Time Domain Measurements for Speech Analysis and Recognition

Abstract

At present, acoustical descriptions of speech sounds are usually given in terms of spectral characteristics, and spectral measurements are the usual form of data input for automatic speech recognizers. Measurements made directly on the acoustic signal are of interest as an alternate form of acoustical description and are potentially very useful in computer-based recognition schemes. A number of such measurements have been investigated. These include zero-crossing rate of the normal and differentiated acoustic signal, time interval between successive zero crossings for normal, low-passed, and high-passed versions of the acoustic signal, form factors of the waveform, and signal intensity Sound classes that have been studied include voiced and unvoiced fricatives, voiced and unvoiced stops, and vowels. Present results indicate that in comparison with spectral measurements: (a) rapid sequential features are better preserved; (b) the measurements are less sensitive to intensity fluctuations; (c) unvoiced fricatives and stops are easily characterized; (d) vowel resolution appears to be comparable. Application of these results to spoken word recognition is considered.