Abstract
Sonagrams of CVC (C = consonant, V = vowel) syllables spoken by a single male speaker, and VC, CV, CVC syllables spoken by two male speakers were converted into 140-dimensional and 144-dimensional quantized patterns respectively. It has been shown that when the components are optimally weighted before being summed, the weighted sum can serve as the basis of recognition. An index of correct recognition has been defined, and a rule for arriving at the correct weights of the individual components has been stated. The sample on which the work has been done is small. Nevertheless, it is of some significance that a procedure of this kind may lead to the identification of vowel classes, syllables and even speakers. The use of adaptive pattern recognition technique is suggested for the recognition of the weighted patterns and an automatic weight determining device is proposed to optimize the weights during the learning phase of the machine.