BIOACOUSTIC SEGMENTATION BY HIERARCHICAL DIRICHLET PROCESS HIDDEN MARKOV MODEL

Vincent Roger, Marius Bartcus, Faicel Chamroukhi, Hervé Glotin

Abstract: Understanding the communication between different animals by analysing their acoustic signals is an important topic in bioacoustics. It can be a powerful tool for the preservation of ecological diversity. We investigate probabilistic models to analyse signals issued from real-world bioacoustic sound scenes. We study a Bayesian non-parametric sequential models based on Hierarchical Dirichlet Process Hidden Markov Models (HDP-HMM). The model is able to infer hidden states, that are referred here as song units. However, using such a model raise one main issue: defining the number of hidden states the model has to learn. In bioacoustic problems we often do not know the number of song units (unlike in human speech recognition). Hence, we work with the Hierarchical Dirichlet Process (HDP)-HMM, which is a Bayesian non-parametric (BNP) model that offers a way to tackle this challenging problem. We focus our work on unsupervised learning from bioacoustic data. It consists in simultaneously finding the structure of hidden song units and automatically infer the unknown number of the hidden states to represent the data. Two real bioacoustic sound scene applications are investigated in this work: on whale and multi-species birds segmentation. The learning of these models is proceeded by using Markov-Chain Monte Carlo (MCMC) sampling techniques on Mel Frequency Cepstral Coefficients (MFCC) of audio signals. The results show an interesting song unit segmentation of the bioacoustic signals and open new insights for unsupervised analysis of such signals. This paper illustrates the potential of chunking non-human animal signals into structured parts. This can yield to a new species representation and help experts to better understand the behaviour of such species as Kershenbaum et al. (2014) wanted.

Conflicts: no conflict : common direction of Phd thesis