Fast Implementation of a Bayesian Unsupervised Segmentation Algorithm

TL;DR

This paper presents a fast, parallel implementation of a Bayesian unsupervised audio segmentation algorithm, improving computational efficiency and analyzing parameter sensitivity with real-world underwater signals.

Contribution

The paper introduces a highly optimized, parallelized version of the Bayesian segmentation algorithm using Cython, GSL, and OpenMP, enabling practical application to real-life signals.

Findings

Significant speed-up in algorithm performance.

Robustness to parameter variations demonstrated.

Effective segmentation of underwater audio signals.

Abstract

In a recent paper, we have proposed an unsupervised algorithm for audio signal segmentation entirely based on Bayesian methods. In its first implementation, however, the method showed poor computational performance. In this paper we address this question by describing a fast parallel implementation using the Cython library for Python; we use open GSL methods for standard mathematical functions, and the OpenMP framework for parallelization. We also offer a detailed analysis on the sensibility of the algorithm to its different parameters, and show its application to real-life subacquatic signals obtained off the brazilian South coast. Our code and data are available freely on github.