Digital Signal Processing and System Theory

Talk Martin Nahrwold

Species/Speaker Recognition with Gaussian Mixture Models - Model Based Marine Mammal Classification -

Date: 10.11.2011, 13:00 h - 14:00 h, Room: Chair library

Martin Nahrwold
CAU, Kiel, Germany


Sunlight rapidly fades beneath the surface of the oceans. Below 1000m the ocean is a place of complete darkness. In the underwater world sound is king. Marine mammals use sound to naviga- te and to detect predators and prey. They use echolocation to obtain environmental information such as water depth, the location of food, or the distance of objects. Ship noise and the use of active sonar systems has an impact on this ability of marine mam- mals. Navies employing high intensity sonar use crew members as marine mammal observers and passive acoustic monitoring as mitigation measures. But marine mammals spend very little time on the surface and hence they are not detectable on a visual survey. So the only possible solution is a passive monitoring system. For the human ear it is a quite easy task to differentiate between species. Like we differentiate between individual speakers just by their voice. Since spectrogram of speech signals on one hand and marine mammals signals on the other hand look rather similar (see figure below), methods known from speaker recognition should be adapted to differentiate between several mammal species.
During this thesis a speaker/species recognizer should be developed. Therefore, different me- thods for feature extraction and model generation should be implemented in Matlab. Some of these features are mel frequency cepstral coefficients (MFCCs) or linear predictor coefficients (LPCs). The features should be used to build up Gaussian mixture models (GMMs) for diffe- rent marine species on one hand and different speakers on the other hand. The results of the individual methods should be compared in terms of their applicability for speaker and species recognition. The recognition scheme with the best common performance should be implemented as a plugin for the DSS real-time framework using C.