Gaussian mixture model classification of frogs

Abstract

This Work Deals With Automatic Call-Independent Frog Species Identification. An Algorithm Is Designed To Process Field Recordings And Perform Automatic Identification Of 10 Species Of Anurans Inhabiting The Yasuní National Park In The Ecuadorian Amazon Region. First, End-Point Detection Using Short-Term-Energy (STE) With A Moving-Average Filter Is Applied To Isolate Frog Calls Over An SNR > 15 Db Threshold. Audio Segments With Background Noise And Silence Are Discarded. Isolated Segments Are Then Parametrized Using Cepstral Feature Vectors That Represent The Frog Acoustic Phenomenon. The Data Is Divided Into Two Groups From Which One Is Used To Train Gaussian Mixture Models And The Others Are Used For Testing Classification Accuracy For Each Species. GMM Models With Different Mixture Weights (Components) Are Generated In Order To Determine The Best Model Order. The Classification Task Is Based On The Maximum-Likelihood (ML) Rule Achieving The Maximum Average Success Rate Of 97.24% With GMM Models Of 64 Components.