Development of an Automatic Segmentation Model for Quran Verses at phoneme level

Abstract

In this research, two models were built to obtain automatic segmentation of the verses of the Holy Quran at the phoneme level, in the first model the Hidden Markov Model Tool Kit HTK was used, the second model was built using the KALDI toolkit at phoneme. to obtain the objectives of this research two data sets have been used. The first data set consisted of a database of 2 hours with the voice of 10 Sudanese male reciters who recited the Quran carefully under the supervision of an expert in the correct and most accurate recitations of the ascription to the Messenger (Peace Be Upon Him), the recordings of 16 Surahs of the Quran. the recitation commenced orderly as in the Holy Quran, starting from Surah Al-Bayyinah to Surah An-Nas. the second data set has been recorded or 100 reciters non-Arab reciters (Indian male) and a total speech corpus of 80 hours in the correct and most accurate recitations. The data set contains surahs (Al-Fatiha, Al-Asr, Al-Kawthar, Al-Ikhlas, Al-Falaq, and An-Nas) the dataset also contains 10 letters which have similar pronunciation. (ظ، ذ، ط، ت، ض، ع، ح، خ، ص، غ) in this research four experiments were evaluated as follows. In the first experiment a process of training and testing the first model was conducted using the first database, the results obtained from the automatic segmentation was 62%. in the second experiment a process of training and testing the second model was conducted using the first database, the results obtained from the automatic segmentation was 62% for test set 70%, and 75% for dev set. in the third experiment a process of training and testing the second model was conducted using the second database but the 10 letters have been omitted, the results obtained from the automatic segmentation was 95% for test set, and 99% for dev set. in the fourth experiment a process of training and testing the second model was conducted using the second database but the 10 letters have been applied, the results obtained from the automatic segmentation was 99.9% for test set and dev set.