Poetic Meter Classification Using Acoustic Cues

Poems, which communicate through rhythm and its apparent meaning have a vital role in any literary. Meter, a set of well defined rules gives rhythm to the poetry. In this paper, a meter classification scheme using fusion of low level, mid-level and high-level musical texture features, computed from recited poems is addressed. The performance of the proposed system is evaluated using a newly created poetic corpus in Malayalam, one of the classical languages in India. Initially, a baseline system with mel-frequency cepstral coefficient (MFCC) feature set is performed. In the second phase, experiment is conducted with musical texture features. In the third phase, experiment is extended using early fusion of MFCC with the feature set considered in the second phase. Support vector based classifier is used in the classification phase. Later, the same feature-sets are experimented with deep neural network(DNN) based classifier. Whilst MFCC-SVM system reports an overall accuracy of 60%, the second phase reported an accuracy of 68%. In the third phase, complementary information provided by the MFCC and musical texture features aided to improve the system performance (accuracy, 90%). In the DNN based experiments, the highest accuracy of 92% is reported for feature-fusion. The experimental study shows the promise of early fusion of MFCC with musical texture feature set in poetic meter classification and its analysis.