Thesis Open Access
Fonseca, Eduardo; Font, Frederic
The goals of this project are the creation of a new dataset of sounds that belong to the domestic environment, called DomesticFSD2018, and to research on methods for the automatic classification of them. A Semi-Supervised approach is used to evaluate the possibility of exploiting samples that are not manually-verified. The purpose of this is to avoid the need of experts and save as many resources as possible in the validation process, that usually takes a lot of time and energies. The train set of DomesticFSD2018 is composed of a trustable (manually-verified) portion of data and a non-trustable (which has received no human validation and can be potentially inaccurate or mislabeled) one. A purely supervised learning approach is firstly followed, training models with only the trustable portion, and both trustable and non-trustable portions of data. Then the semi-supervised learning approach is experimented, using the models trained in the previous step to make predictions on non-trustable data. The samples predicted with the highest level of confidence are added to the train set, and finally, the classifier is re-trained using the updated and larger train set. In both cases, the technologies used are Support Vector Machines using MFCCs’ properties as input. The semi-supervised approach shows better results and allows us to add a considerable amount of non-trustable data to the trustable portion of the dataset.