Evaluating multiple viewpoint models of tabla sequences
Parag Chordia, Avinash Sastry, Aaron Albin · 2010
We describe a realtime tabla generation system based on a variable-length n-gram model trained on a large symbolic tabla database. A novel, parametric smoothing algorithm based on a family of exponential curves is introduced to control the relative weight of high- and low-order models. This technique is shown to lead to improvements over a back-off smoothing for our tabla database. We find that cross-entropy is lowest when the coefficient of the exponential curve is between 1 and 2 and increases for values outside of this optimal range. The basic n-gram model is extended to model dependencies between duration, stroke-type, and meter using cross-products in a Multiple Viewpoints (MV) framework, leading to improvements in most cases when compared with independent stroke and duration models.