Gamaka Synthesis for Kalpitha Swaras in Carnatic Music

Gamaka (note ornamentation) is an essential element of Carnatic music. Earlier works in computer-generated gamakas focused on developing rule-based mathematical models of each gamaka, which fails to capture the intricate changes in pitch and thus does not sound natural. To address this challenge, this work approaches the synthesis of gamaka for kalpitha swaras (composed notes) in Carnatic music using a data-driven system. The model uses masked latent space representation in an auto-encoder architecture with features extracted using convolutional layers. Gamaka information is embedded and combined with pitch contours extracted from symbolic data. Our model is successful in synthesizing gamaka with nuances that closely follow the ground truth. We evaluated our system on various aspects of computer-generated gamakas, including performance scoring on unseen data, Turing tests, new phrases and ragas, and customizable and tempo-dependent gamakas. Results of the expert listening test have shown a 26% improvement (rel.) on the Mean Opinion Score (MOS), authenticity, and intonation of gamakas for our approach - (MOS 3.91 out of 5) compared to the baseline (Gaayaka Software, MOS 3.1).