Speaker-dependent 1000 word recognition using a large scale neural network 'CombNET-II' and dynamic spectral features

The authors describe speaker-dependent large vocabulary word recognition using a large-scale neural network, CombNET-II, 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-II consists of two types of neural networks. The first part is a stem network which learns by a self-growing algorithm and roughly classifies an input pattern. The second part consists of many branch networks which learn by a backpropagation algorithm and precisely classify the input pattern. A stem network is a vector quantizing network and it reduces the number of category candidates for the branch networks, so that each branch network has only a small number of connections and it is easy to tune up. Experiments on speaker-dependent large-vocabulary word recognition for 1000 Chinese spoken words is described. Experimental results show that the high recognition accuracy of 99.1% is obtained and that CombNET-II is very effective for large vocabulary spoken word recognition.<>