A new parameter smoothing method in the hybrid TDNN/HMM architecture for speech recognition

Abstract

In this paper, we propose a new parameter smoothing method in the hybrid time-delay neural network (TDNN)/hidden Markov model (HMM) architecture for speech recognition. In the hybrid architecture, the TDNN and the HMM are combined using the activations from the second hidden layer of TDNN as the outputs of a fuzzy vector quantizer (FVQ). The HMM algorithm is modified to accommodate these FVQ outputs. in our modular construction of TDNN, the input layer is divided into two states to deal with the temporal structure of phonemic features, and the second hidden layer consists of two states in a time sequence. To improve the performance of the hybrid architecture, a new smoothing method has been proposed. The average values of the activation vectors from the second hidden layer of the modular TDNN are used to generate the smoothing matrix from which smoothed output symbol observation probability is obtained. With this proposed approach, our simulation results performed on speaker-independent Korean isolated words show the reduction of the error rate by 44.9% as compared to the floor smoothing method.