ETHIOPIAN ORTHODOX TEWAHEDO CHURCH ZEMA GENRE CLASSIFICATION USING DEEP LEARNING A Thesis Submitted to the School of Graduate Studies in Partial Fulfillment of the Requirement for the Degree of Master of Science iInformation Technologyn

Abstract

Research has been conducted to automate the classification of music genres based on their sound, with applications in music information retrieval (MIR) systems, music data analysis, and music transcription. As the volume of music data continues to grow, indexing and retrieval pose significant challenges. Previous studies have focused on classifying, identifying, predicting, and distinguishing music genres for contemporary music services. However, the absence of a standardized classification system for Zema genres hinders the successful preservation, comprehension, and dissemination of this rich musical tradition. This study specifically addresses the classification of Ethiopian Orthodox Tewahedo Church Zema, which falls within the domain of music information retrieval. Geez Zema, a traditional form of education in the Ethiopian Orthodox Tewahedo Church (EOTC), involves priests performing measured sounds while dressed in secular attire. Due to its inherent connection to music and rhythm, it can be considered a form of secular art. The primary motivation for this research stems from the knowledge gap between modern and traditional education regarding the Zema genre, as many students in this traditional school (Abinet School) lack a comprehensive understanding of it. Therefore, a model was developed to categorize the sound of Geez Zema into its respective genre, aiming to address this limitation. Three major Zema types, namely Geez (ግዕዝ), Ezil (ዕዝል), and Araray (ዓራራይ), were utilized to classify Geez Zema. Audio data was collected from the traditional religious school (Abinet School) and recorded using smartphones and the different school's websites for teaching purposes. Following data collection, the audio underwent preprocessing, including windowing with predetermined durations of 10 seconds, 15 seconds, and 20 seconds. Then feature selection and extraction were held and we selected nine different features. Subsequently, a model was created by extracting features and employing a deep learning approach for classification. Specifically, convolutional neural network (CNN), recurrent neural network (RNN), and convolutional recurrent neural network (CRNN) models were developed using the softmax classifier. The CRNN model with a window size of 20 seconds and a 90/10 train-test split achieved training accuracy of 98.9% and test accuracy of 97.89%, precision of 98 %, recall of 98%, and F1-score 98%.