Automatic Arabic Speech Recognition by CNN

Abstract

Automatic speech recognition has been an active field of study since the 1950s, which explains its richness but also its difficulty. It involves the collaboration of multiple disciplines and techniques. The complexity of the speech signal, resulting from the interaction between sound production and perception by the ear, contributes to the challenge of automatic speech recognition, which has become a highly interesting research topic. The objective of this thesis is the acquisition and implementation of a database consisting of 20 phrases divided in to two categories connected words and separated words, along with a corpus of syntactically and semantically correct sentences. This database was recorded under real conditions, and the acoustic analysis of this database was performed using the MFCC method, providing us with a series of input vectors for the implemented Automatic Speech Recognition (ASR) system. This system is based on Convolutional Neural Networks. Evaluating the performance of the ASR system using the database analysis method will highlight the influence of parameterization.