A method of analysis and recognition for voiced vowels

A method of speech analysis that has been shown to be capable of recognizing with high accuracy a set of seven voiced vowels spoken by twelve male talkers with various regional accents is described. The waveforms used in the recognition scheme are tapped from four points along a low Q dispersive delay line, which represents a model of the human cochlea. These four output signals are sampled for 4 ms at a rate of 25 000 points per second per output channel. sampling time is synchronized withthe onset of a glottal pulse. The 4-ms samples are autocorrelated on a digital computer, then crosscorrelated (at zero delay only) with a set of stored prototype patterns to produce an array of cross-correlation coefficients. These coefficients are treated as components of a multidimensional vector that characterizes the input sound. The final decision as to which sound was spoken is made by a simple linear adaptive network that was trained to seperate these multidimensional vectors into their proper classes. The network repeatedly alters a set of decision surfaces until correct classification has been obtained or until a specified number of trials has been exceeded. Successful training was attained in all cases indicating a linear separabilitly of the vowel sounds in the space described by the correlation operations, and the short sampling time used points up the desirability of short time-signal analysis techniques in speech recognition work.