Speaker-dependent 100 word recognition using CombNET and dynamic spectral features of speech

Abstract

Present speaker-dependent 100-word recognition using CombNET, which consists of a four-layered neural network with a comb structure, and dynamic spectral features of speech based on a two-dimensional mel-cepstrum. CombNET consists of two types of neural network. The first one is a stem network which utilizes a self-organizing algorithm and roughly classifies an input pattern. The second one consists of many branch networks using a back-propagation algorithm and precisely classifies the pattern. Experimental results on speaker-dependent word recognition for 100 Japanese city names uttered by nine male speakers show that the recognition accuracy is 97.3%.<>