The Applicability of Miremotion in Emotion Classification of Sri Lankan Folk Melodies
Posted: 2 Dec 2019
Date Written: December 2, 2019
Abstract
Music evokes powerful emotions, and this ability has motivated researchers worldwide to explore relationships between music and emotion. Even though numerous classification tasks have been carried out, frequently reported discrepancies among classification performances could be partially attributed by the introduction of independent datasets, classification models, and different study populations. While studies in music psychology provide exciting evidence for universality in music-emotion perception, various scholars have emphasized the impact of socio-cultural background and the individual differences of listeners especially in perception of more complex emotions. Given this background, we in a preliminary study, tested the generalizability of miremotion a classification model (trained using western film music) defined in Matlab MIRToolBox. The model was tested using Sri Lankan folk music (30 stimuli; 30s; 44100Hz; stereo; 32bit; .wav) representing happy, sad, and fear. The above melodies were chosen as they are abundant in emotion expressivity yet remain to be computationally analysed. For each stimulus, predicted scores for happy, sad, tender, anger, and fear were obtained using miremotion. A listening experiment was conducted (Sri Lankan undergraduates; 25-30yrs; no specialized music education), through which the ratings for same emotion categories were obtained in a seven-point Likert scale. Paired t-tests were performed to compare objective predictions and subjective ratings. Other than for ‘anger’ (p > 0.05), all other emotions showed significant differences (‘happiness’ (p < 0.001), ‘sadness’ (p < 0.05), ‘tender’ (p < 0.05), ‘fear’ (p < 0.05)). Findings reveal that the generalizability of existing classifiers depends on the emotion category and the study further supports the need for developing own classifiers, especially for culturally-specific music. However, the study was limited by the inclusion of a limited data set and a homogeneous study population. Extensions in the above regards would uncover compelling findings on universality and cultural specificity in music-emotion and their acoustic correlates.
Keywords: music emotion classification, matlab MIRToolbox, miremotion, paired t-test, Sri Lankan folk melodies
Suggested Citation: Suggested Citation