An Experimental Speaker-independent System for Isolated Word Recognition Implemented for Romanian Language

The research presented in this paper is based on the artificial neural networks recognition paradigm applied to Romanian isolated word recognition. The network, which is composed by three layer (a Multilayer Perceptron), is trained by conventional Back-propagation algorithm. The ANN speech recognition system based on Mel Frequency Cepstral Coefficients was developed using Matlab toolkit. The system was tested on the 60 isolated word Romanian speech corpus and it was training on a 60 distinct isolated word recognition task. The word recognition accuracy obtained was about 81.6 %. Here we presented a set of experiments involving artificial neural networks used in a speech recognizer system. The goal was to test our program for different speakers and to optimize a number of factors like: structures of the network, input window size, network training procedure, etc. All this factors can influence the performance of our system. The platform for research and development of our recognition system is using a program that was developed in Matlab. 2000 Mathematics Subject Classification: 68T10, 82C32, 62M45. 1. Implementation of an artificial neural network for speech recognition The paper presents an application that implements a type of the artificial neural network: MLP (Multi Layer Perceptron). This network is based on Backpropagation learn, which is the adjustment of the weights according to the global error of the network. This adjustment is done in backward order of processing.