The research topics I am currently working on are:
1. The first step is to reduce the noise effects in either time or spectral domain depending upon the application. I am developing practical algorithms to detect dolphin whistles from underwater sounds recorded in the ocean or captive conditions. This goal can be achieved both in time domain by designing special filters such as energy detector, etc. and in spectral domain by making the spectrogram treated as an image and then applying image-processing techniques to spot the whistle such as spectral correlation. The followings are typical spectrograms of bottlenose dolphin whistles as part of our current database collected by Woods Hole Oceanographic Institution (WHOI):
2. Another issue which needs to be addressed is feature extraction. Some methods require tracing the whistle contour in order to extract the important features such as minimum and maximum frequencies, temporal duration, number of inflection points, beginning and ending slopes, etc while some other image-processing based approached don't and recognize the important patterns within the image reducing costly computations such as Local Binary Patterns (LBP), Grid Masking, or other different methods like modelling of Gabor Wavelets in order to obtain the important information of the image so that there is no need to extract the contour.
3. The classification can be performed to distinguish different types of calls or various species based on the characteristics of their vocalizations. The approaches developed in this research are:
- K-Nearest Neighbor (KNN):
- Support Vector Machine (SVM): both linear and nonlinar
- Compressive Sensing classifier: based on Sparsity criteria
- Hidden Markov Model (HMM)