Oktoechos Classification and Generation of Liturgical Music using Deep Learning Frameworks

Abstract

An important feature of the music repertoire of the Syrian tradition is the system of classifying melodies into eight tunes,  called ’oktoe\={c}hos’.  In oktoe\={c}hos tradition, liturgical hymns are sung in eight modes or eight colours (known as eight ’niram’ in Indian tradition). In this paper, recurrent neural network (RNN) models are  used for  oktoe\={c}hos genre classification with the help of musical texture features (MTF) and i-vectors.The performance of the proposed approaches is evaluated using a newly created corpus of liturgical music in the South Indian language, Malayalam. Long short-term memory (LSTM)-based and gated recurrent unit(GRU)-based experiments report the average classification accuracy of  83.76\%  and 77.77\%, respectively, with a significant margin over the i-vector-DNN framework.   The experiments demonstrate the potential of RNN models in learning temporal information through MTF in recognizing eight modes of oktoe\={c}hos system. Furthermore, since the Greek liturgy and Gregorian chant also share similar musical traits with Syrian tradition, the musicological insights observed can potentially be applied to those traditions. Generation of oktoe\={c}hos genre music style has also been discussed using an encoder-decoder framework. The quality of the generated files is evaluated using a  perception test.