SPEECH-RECOGNITION OF ISOLATED ARABIC WORDS USING ARTIFICIAL NEURAL NETWORKS

Abstract

Speech recognition belongs to the broader category of pattern recognition tasks performed by many Neural Network models . This thesis investigates different supervised and unsupervised learning techniques for speech recognition . Specifically , KOHONEN self-organizing feature map (KSFM) and the differential competitive learning (DCL) as unsupervised techniques ; vector quantization (VQ) methods, CHAN & WU density estimator (CWDE) and the backpropagation (BACKP) learning algorithm as supervised ones .

The thesis concentrates on the speaker-independent isolated-words recognition case , so one is faced with the need of an accurate segmentation and endpointing of the words which is a major requirement of the Neural Network models.

A test problem differentiating the Arabic words ('...//'..:.:!' ') was investigated

, automatically endpointing the utterances around manually endpointed templates by applying dynamic time warping (DTW) . However .. this way of analysis is only useful to compare the investigated Neural Network models, since it is used when testing is done offline , with the utterances' classes known in advance .

Trying to build a practical online isolated-words recognition system , another model where accurate segmentation is not a must, is demanding .The time-delay Neural Network (TDNN) is a hierarchical model , based on the Backpropagation learning algorithm , that was found able to learn precise weight patterns from unprecisely prepared training examples .