A Framework for Bioacoustic Vocalization Analysis Using Hidden Markov Models

Abstract

Using Hidden Markov Models (HMMs) as a recognition framework for automatic classification of animal vocalizations has a number of benefits, including the ability to handle duration variability through nonlinear time alignment, the ability to incorporate complex language or recognition constraints, and easy extendibility to continuous recognition and detection domains. In this work, we apply HMMs to several different species and bioacoustic tasks using generalized spectral features that can be easily adjusted across species and HMM network topologies suited to each task. This experimental work includes a simple call type classification task using one HMM per vocalization for repertoire analysis of Asian elephants, a language-constrained song recognition task using syllable models as base units for ortolan bunting vocalizations, and a stress stimulus differentiation task in poultry vocalizations using a non-sequential model via a one-state HMM with Gaussian mixtures. Results show strong performance across all tasks and illustrate the flexibility of the HMM framework for a variety of species, vocalization types, and analysis tasks

Social Vocalizations of Big Brown Bats Vary with Behavioral Context

“…A variety of techniques exist for automated classification, including frequency contour comparison [45], neural networks [46], hidden Markov models [47], and spectrogram correlation [44], [48], [49]. Although we experimented with other techniques, we chose spectrogram correlation because it performs well, errors are easier to understand, and large training data sets are not required [44].…”

Automated classification of mouse pup isolation syllables: from cluster analysis to an Excel-based “mouse pup syllable classification calculator”

“…Even so, it allows the reader to better understand the features used in the classification, as it is not always clear what features are used in subjective categorizations and how distinct these features are. While other automated call classification methods have been described, such as hidden Markov models (Ren et al, 2009) and neural network models (Pozzi et al, 2010), these depend on pre-determined categories.…”

“…Several techniques have been implemented for the automatic identification of syllable types in other species, including hidden Markov models (Ren et al, 2009), neural network models (Pozzi et al, 2010), and forms of cluster analysis (Takahashi et al, 2010). Hammerschmidt et al (2012) used two-step cluster analysis to classify infant mouse vocalizations into two categories, short syllables and long syllables.…”

Discrimination of individual tigers (Panthera tigris) from long distance roars

“…Twelve mean-normalized GFCC coefficients plus normalized log energy, along with velocity and acceleration, are computed for a total of 39 features per frame (Ren et al, 2009). The programming toolkit used to implement feature extraction, as well as HMM training and testing, is the hidden Markov model toolkit from Cambridge University (Young et al, 2006).…”

“…Greenwood frequency cepstral coefficients (GFCCs) (Clemins and Johnson, 2006;Ren et al, 2009) were used as the primary frame-based spectral features for the HMMbased classification experiments. GFCCs are a generalization of the Mel-frequency cepstral coefficient representation which is widely used in human speech processing and recognition, adapted to characterize a broader range of species.…”