A Parametric Method for Multi-Pitch Estimation

Abstract: This thesis proposes a novel method for multi-pitch estimation. The method operates by posing pitch estimation as a sparse recovery problem which is solved using convex optimization techniques. In that respect, it is an extension of an earlier presented estimation method based on the group-LASSO. However, by introducing an adaptive total variation penalty, the proposed method requires fewer user supplied parameters, thereby simplifying the estimation procedure. The method is shown to have comparable to superior performance in low noise environments when compared to three standard multi-pitch estimation methods as well as the predecessor method. Also presented is a scheme for automatic selection of the regularization parameters, thereby making the method more user friendly. Used together with this scheme, the proposed method is shown to yield accurate, although not statistically efficent, pitch Estimates when evaluated on synthetic speech data.