Though Arabic language is a widely spoken language, research done in the area of Arabic Speech Recognition is limited when compared to other similar languages. Also, while the accuracy of speaker dependent speech recognizers has nearly reached to 100%, the accuracy of speaker independent speech recognition systems is still relatively poor.
This paper concerns with the recognition of speaker independent Arabic speech using Support Vector Machine. The proposed model is applied on the connected Arabic digits (number) using Neural Networks as an example. Also we can apply the system to any other domain.
A spoken digit recognition process is needed in many applications that use numbers as input such as telephone dialing using speech, airline reservation, and automatic direc-tory to retrieve or send information.
This has been realized by first building a corpus consisting of 1000 numbers compos-ing 10000 digits recorded by 20 speakers different in gender, age, physical conditions…, in a noisy environment. Secondly, each recorded number has been digitized into 10 sepa-rate digits. Finally these digits have been used to extract their features using the Mel Frequency Cepstral Coefficients (MFCC) technique which are taken as input data to the Neural Networks for the recognition phase.
The performance of the system is nearly 94% when we used the Support Vector Ma-chine (SVM). |
