NL7400088A - - Google Patents

Abstract

1433770 Speech recognition systems XEROX CORP 19 Dec 1973 [5 Jan 1973] 58825/73 Heading G4R A speech signal, e.g. a word, is divided into n frequency bands, and in each of a number of successive time intervals the energy level in each band is used to define a set of n values which may be regarded as defining a point in n dimensional space. Thus the successive sets represent successive points on a curve in that space. Each set (point) is inspected to determine which of a number of reference sets (points) corresponds to that set and is assigned a 'label' representing that reference set. The resulting sequence of labels is compared with reference sequences derived from reference speech signals in order to identify the unknown speech signal with one of the reference signals. As disclosed the speech signal is divided into 6 bands by filters which feed an integrator. The integrator sums the energy in each band over successive periods of 10 m.sec. and feeds a processor via an analog-digital converter. The converter output represents the energy value of each of the 6 banks in each 10 m.sec. interval. Each set of 6 values defines a point in 6-dimensional space and the succession of points a curve, e.g. as the curve shown in Fig. 3 for a 3-dimensional space defined by frequency bands X, Y, Z. The processor normalizes the sets and determines which point of a number of reference points, e.g. p1-p8 in Fig. 3, is nearest to each of the sets of values, and labels the set with the label of that nearest point. In general the number of bits required to define a label is less than the number required to define a set of the values. The sequence of labels is compared with reference sequences as indicated above. Because speakers do not always sound the final stop consonants these consonants are ignored if they last less than 200 m.sec., so that their presence or absence is immaterial to the recognition process.