A recursive Bayesian algorithm for detection of changepoints in unidimensional signals

The problem of detecting changepoints in time series has been studied since at least the 1950s, and has applications in several areas. In this talk we present a brief historical survey of the problem and solutions proposed in the literature. We then propose a recursive algorithm for audio segmentation based on the search of changepoints in the total signal power. The algorithm uses a fully-Bayesian hypothesis test as stopping condition, and has worst-case complexity O(n log n); the operating characteristics of the algorithm can be effectively adjusted based on a single free parameter. We present a Python+Cython implementation, and show an application to the unsupervised analysis of underwater audio signals of long duration.