WASIS - Bioacoustic Species Identication based on Multiple Feature Extraction and Classication Algorithms

Automatic identication of animal species based on their sounds is one of the means to conduct research in bioacoustics. This research domain provides, for instance, ways to monitor rare and endangered species, to analyze changes in ecological communities, or ways to study the social meaning of animal calls in their behavioral contexts. Identi cation mechanisms are typically executed in two stages: feature extraction and classi cation. Both stages present challenges, in computer science and in bioacoustics. The choice of e ective feature extraction and classi cation algorithms is a challenge on any audio recognition system, especially in bioacoustics. Considering the wide variety of animal groups studied, algorithms are tailored to speci c groups. Audio classi cation techniques are also sensitive to the extracted features, and conditions surrounding the recordings. As a results, most bioacoustic softwares are not extensible, therefore limiting the kinds of recognition experiments that can be conducted. Given this scenario, this dissertation proposes a software architecture that allows multiple feature extraction, feature fusion and classi cation algorithms to support scientists and the general public on the identi cation of animal species through their recorded sounds. This architecture was implemented by the WASIS software, freely available on the Web. Since WASIS is open-source and expansible, experts can perform experiments with many combinations of pairs descriptor-classi er to choose the most appropriate ones for the identi cation of given animal sub-groups. A number of algorithms were implemented, serving as the basis for a comparative study that recommends sets of feature extraction and classi cation algorithms for three animal groups.