Sachin Chachada, a MCL member, has passed his defense on Jan 14, 2016. Congratulations!
His dissertation title is “Classification and Retrieval of Environmental Sounds”. Speech and music signals have been extensively studied for several decades. Need for Environmental Sound Recognition (ESR) system has picked up the pace in recent years. Environmental sounds are quotidian sounds, both natural and artificial, i.e. sounds one encounters in daily life other than speech and music. In this thesis, Narrow Band Time Frequency features are proposed which characterize a signal using TF representation of its band limited filtered signal. In order to further improve the performance, use of a novel multi-classifier approach, Para-Boost (PB) model, is proposed. It takes the advantages of all the features and improves the overall performance of ESR system. Finally, considering the exponential growing environmental sound data on the Internet, the thesis tackles the problem of a good content based retrieval system. A two stage content based environmental sound retrieval system is proposed. This query-by-example retrieval system assumes that the database is partially labeled. In Stage I, a broad categorization of environmental sounds based on their signal-characteristic is done. For each category, a classifier is trained to predict labels for unlabeled data in the database and also narrow search range for a query by assigning it multiple, yet limited, class labels. In Stage II, a novel feature and a scoring scheme to do local matching and ranking is proposed. An audio signal is first segmented in an unsupervised manner using Mean Shift algorithm and each segment is represented by its point of convergence in the feature space. The audio signal is finally represented by its energy distribution over each segment thereby capturing the temporal variations of the audio signal in feature space. Given a query, first audio segments of a document are mapped to those of the query. Then the document is assigned a score based on energy distribution of mapped segments only. The framework performs well even with 90% unlabeled data.
Sachin gave a clear and smooth talk. The committee was impressed with the research work and gave valuable suggestions. When talking about his success in his research work, Sachin shared his experience with us. He believed his journey at MCL group had taught him a lot more than the art of research. The group fosters a culture of sharing and helping which, he believes, got him through his PhD. He considered himself lucky to have had the opportunity to train under Dr. Kuo, who was not only a great advisor for research, but also a life coach. He feels this experience at MCL has rendered him capable of independently working on research problems and is prepared to start the net journey of his life as a research professional.
