Proceedings of meetings on acoustics · 2012 · 11 citations · 0 references
Natural SciencesMarine SciencesZoologyMarine MammalogyMarine ZoologyUnderwater AcousticsBird VocalizationHidden Markov ModelsMarine Mammal AcousticsMarine Mammal BioacousticsPattern RecognitionSpeech RecognitionCall RecognitionCetacean VocalizationsClassification AlgorithmsSupervised ClassificationSupervised Classification AlgorithmsTrainingWhale BioacousticsWhale Song
Hidden Markov Models (HMMs) are widely used for speech recognition tasks but have been used rarely for the classification of bioacoustics signals. This study proposes a new approach for the supervised classification of the calls that compose humpback whale (Megaptera novaeangliae) song sequences with the use of HMMs, based on the concept of subunits as building blocks. HMMs are suitable candidates to conduct such classification tasks because they are very flexible, which suggests that they will cope with the great variability of the humpback whale call repertoire and also with the changes in duration of individual calls. Indeed, the same vocalisations which are repeated by an individual throughout a song may vary in length as the song progresses, posing a challenge for classification algorithms. Another attractive characteristic of HMMs is that highly developed tool-set is widely available. We describe the HMM classification method and show that a high level of performance can be achieved with modest requirements both in terms of computational load and storage. Training stage requires minimal manual input and once trained the recognition process is fully automated. We will present how the classification performance of songs recorded in Madagascar is affected by different amount of training.