Adventures in Carnatic music percussion modeling: A progress report (presented by Konstantinos Trochidis)
In this presentation we will introduce a software prototype in Max for automatic Carnatic music percussion generation in Adi tala. The algorithm for Carnatic music percussion generation follows a data-driven approach in which several hours of percussive sequences performed by Carnatic percussionist Akshay Anantapadmanabhan were recorded. All recordings were manually annotated using Sonic Visualizer with beats and downbeats. Each stroke event was coded as a string based on its frequency content (Lo-Mid-Hi), the inter-onset-interval (IOI) duration of the stroke and a value indicating the velocity by computing an onset detection function and estimating its amplitude level.
The sequences were subsequently parsed using Godfried Toussaint’s Mutual Nearest Neighbor grouping algorithm (Toussaint 2016). We used the bags of words approach which is commonly used in document classification and clustering as a tool of feature generation. We transformed the text of the groupings to a “bag of words” vector representing the frequency of occurrence of each unigram term in the groupings. This leaded to the generation of a feature matrix for all the groupings, which further used for clustering analysis in terms of similarity using the K-means clustering approach. The clusters were then mapped in 2D space using t-SNE (van der Maaten & Hinton, 2008) visualization technique to convert a high-dimensional data set into a matrix of pairwise similarities, and computed the Euclidean distance between clusters of rhythm patterns. This space, depicted graphically, allows the users to travel though rhythmic sequences that are grouped by similarity, allowing continuous progression and variation according to similarity in the rhythmic space. The model has been developed in straight collaboration with Anantapadmanabhan, who has been guiding the group on how to devise correct rhythmic sequences in Carnatic music.