Phoneme-based recognizer to assist reading the holy Quran

Abstract

This paper presents a new phase of our ongoing efforts for building a high performance speaker independent recognizer for Quran recitation. An inhouse developed and annotated sound database of about eight hours is used for this purpose. Since this sound database is segmented and annotated on both allophone and phoneme levels, we are developing two separate baseline recognizers for respectively allophones and phonemes. We employed the same approach for developing both phoneme and allophone recognizers to be able to make some kind of comparison between them. The Cambridge HTK tools are used for the development of these recognizers. We present in this paper the development of the phoneme-based recognizer to measure its appropriateness for the sake of our ultimate goal of building a high performance speaker independent recognizer to assist reading and memorizing the Holy Quran; the details of the allophonic recognizer is being published separately. Each Quarnic phoneme is modeled by an acoustic Hidden Markov Model (HMM) with 3-emitting states. A continues probability distribution using 16 Gaussian mixture distributions is used for each emitting  state. Results give 92% of average recognition rate, which is very promising, compared to 88% for the allophonic recognizer.