Abstract:
In this study of sound recognition for selected verses of the Holy Quran, five
verses selected from the Holy Quran. This done as a small number of verses contains
all the Arabic phonemes. One hundred recitations of each verse were prepared in
wave files. Verses had been recited by famous certified readers of the Holy Quran. A
file created from each recite by extracting only the first word of the verse. A wave file
of noise is set by the researcher's voice, for testing purpose. All wave files recorded at
Sampling rate =22.050 kHz ,
PCM signed 16 bit mono.
Three types of coefficients are extracted from each wave file to represent
features of speech. They are Mel Frequency Cepstrum Coefficients MFCC, Power
Spectral Density PSD and Reflection Coefficients RC.
Also three techniques of speech recognition are used. They are Hidden
Markov Models, Dynamic time warping and Artificial Neural Networks.
Test were done at two levels. The first stage was applied at the full verse level.
All the three techniques mentioned were used. HMMs and ANNs trained by the first
30 samples and test done by all the 100 sample of each verse.
The second stage was applied in the same way, but the used samples were of
the first word of the verse. HMMs technique only was selected to be used in
recognition. That due to high recognition rates scored at the first stage. This stage
repeated in the same way , but used the first 50 samples in training instead of the first
30 samples.
Mainly, HMMs scored high rates of recognition to coefficients used(MFCC, PSD
and RC). Low recognition rates with high confusion scored by ANNs and DTW at
IIIverse level. For all coefficients used, high scores of recognition rates with low
confusion rate concentrated in HMMs with MFCC. MFCC scored higher recognition
rates than PSD and RC.
