Artificial Neural Network for Arabic Speech Recognition in Humanoid Robotic Systems

Abstract

Speech recognition is projected to play an increasingly important role in the future of human/machine interfacing. The objective of this study is to engineer a speech recognition system capable of deployment onto a general humanoid robot. A MATLAB based program for speech extraction and identification is constructed. Although there are many different algorithms used in speech recognition, the utilization of artificial neural networks (ANNs) was found to be adequate for the Arabic language, with its multitude of complexities, accents and linguistic intentionality. Furthermore, ANN is powerful and can model complex functions while offering the opportunity for additional cognitive abilities. The software tool developed converts the incoming audio signal into a two dimensional spectrogram, which is subsequently supplied to the ANN through a mel frequency cepstral coefficients (mfcc) algorithm. A software product capable of converting Arabic speech to commands was developed for controlling the “NAO” humanoid robot with 90% hit rate.