Semi-Automatic Annotation of Music Collections

The amount of multimedia content in the World Wide Web is increasing very much, and music is one of the most outstanding. Every time, there are more and more songs, artists, and even new genres. Hence, it is really hard to manage this huge quantity, in terms of searching, filtering, navigating through the content, etc. One of the solutions for this problem is keeping annotations of the music files, in order to facilitate the retrieval process. However, it is known that annotating songs manually has a huge cost and annotating them automatically is quite inaccurate. The approach of this master thesis is to propose a semi-automatic strategy that allows to annotate huge music collections, based on audio similarity and a community of users that annotate music titles. This strategy allows to increase the efficiency regarding the manual annotation, and the accuracy regarding the automatic annotation. The Thesis presents two experiments followed for the evaluation of the annotation process: the first experiment consists on testing how the content–based similarity can propagate labels. Using a collection of of ∼5500 songs, we show that with a collection annotated at 40% with styles, we can reach a 78% (40%+38%) annotated collection, with a recall greater than or equal to 0.4, only using content–based similarity. In the case of moods, with a 30% annotated collection we can automatically propagate up to 65% (30%+35%). Regarding the second experiment, we use a collection of ∼258000 songs. With a 48% manually annotated collection we propagate the annotations up to 76% (48%+28%) and then evaluate a small set of the propagated annotations by means of user relevance feedback.