Critiquing Task- versus Goal-oriented Approaches: A Case for Makam Recognition

Computational Musicology and Music Information Retrieval (MIR) address the core musical question under study from a different perspective, often a combination of top-down vs. bottom-up approaches. However, the evaluation metrics for MIR tend to capture the model accuracy in terms of the goal. For instance, mode (melodic framework) recognition is implemented with a goal to evaluate and compare melodic analysis approaches, but it is worth investigating if at all it lends itself as one befitting proxy task. In this work, we aim to review whether the model actually learns the task it is intended for. This is particularly relevant in non-Eurogenetic music repertoires where the grammatical rules are rather prescriptive. We employ methodologies that combine domain-knowledge and data-driven optimizations as a possible way for a comprehensive understanding of these relationships. This is tested on Makam which is one of the understudied corpora in MIR. We evaluate an array of feature-engineering methods on the largest mode recognition dataset curated for Ottoman-Turkish makam music, composed of 1000 recordings in 50 makams. We adapted the time-delayed melody surfaces (TDMS) feature, which in combination with support vector machine (SVM) classifier yields 77.2% recognition accuracy, comparable to the current state-of-the-art. We also address (ethno)musicology-driven tasks with a view to gathering deeper insights into this music, such as tuning, intonation, and melodic similarity. We aim to propose avenues to extend the study to makam characterization over the mere goal of recognizing the mode, to better understand the (dis)similarity space and other plausible musically interesting facets.