Speech recognition using autocorrelation analysis

Abstract

Experiments are described in which word recognition is based on digital autocorrelation analysis followed by computer pattern matching. Incoming speech is split into two frequency bands, and the signals in each band are quantized into two amplitude levels. The two signals are fed to separate autocorrelators, consisting of binary shift registers, digital multipliers, and RC integrators. The low- and high-frequency correlators have, respectively, 10 and 8 outputs which are coded into a 36-bit character, sampled 40 times per second, and fed to a digital computer for recognition. In the computer, master patterns in the form of a 36 ×30 matrix, are generated for each word of the vocabulary from a number of known utterances of the word. Unknown utterances are then compared with each master pattern in turn, and the best match determined by a simple scoring technique; if desired, master patterns can be "updated" when correct recognition occurs. Master patterns can be formed from either one or several speakers; when formed from a single speaker, and with a vocabulary of 10 words, subsequent utterances by the same speaker are recognized with an average accuracy of 90 percent.