Unsupervised Bioacoustic Segmentation by Hierarchical Dirichlet Process Hidden Markov Model.

Bioacoustics is powerful for monitoring biodiversity. We investigate in this paper automatic segmentation model for real-world bioacoustic scenes in order to infer hidden states referred as song units. Nevertheless, the number of these acoustic units is often unknown, unlike in human speech recognition. Hence, we propose a bioacoustic segmentation based on the Hierarchical Dirichlet Process (HDP-HMM), a Bayesian non-parametric (BNP) model to tackle this challenging problem. Hence, we focus our approach on unsupervised learning from bioacoustic sequences. It consists in simultaneously finding the structure of hidden song units, and automatically infers the unknown number of the hidden states. We investigate two real bioacoustic scenes: whale, and multi-species birds songs. We learn the models using Markov-Chain Monte Carlo (MCMC) sampling techniques on Mel Frequency Cepstral Coefficients (MFCC). Our results, scored by bioacoustic expert, show that the model generates correct song unit segmentation. This study demonstrates new insights for unsupervised analysis of complex soundscapes and illustrates their potential of chunking non-human animal signals into structured units. This can yield to new representations of the calls of a target species, but also to the structuration of inter-species calls. It gives to experts a tracktable approach for efficient bioacoustic research as requested in Kershenbaum et al. (Biol Rev 91(1):13-52, 2016).