Automatic Identification of Arabic Dialects USING Hidden Markov Models
Fawzi Suliman Alorifi · D-Scholarship@Pitt (University of Pittsburgh) · 2008
The Arabic language has many different dialects, they must beidentified before Automatic Speech Recognition can take place.This thesis examines the difficult task of properly identifyingvarious Arabic dialects. We also present a novel design of anArabic dialect identification system using Hidden Markov Models(HMM). Due to the similarities and the differences between Arabicdialects, we build a ergodic HMM that has two types of states; oneof them represents the common sounds across Arabic dialects, whilethe other represents the unique sounds of the specific dialect. Wetie the common states across all models since they share the samesounds. We focus only on two major dialects: Egyptian and theGulf. An improved initialization process is used to achieve betterArabic dialect identification. Moreover, we utilize many differentcombinations of speech features related to MFCC such as timederivatives, energy, and the Shifted Delta Cepstra in training andtesting the system. We present a detailed comparison of theperformance of our Arabic dialect identification system using thedifferent combinations. The best result of the Arabic dialectidentification system is 96.67\% correct identification.