Automatic Segmentation for Music Classification using Competitive Hidden Markov Models

TitleAutomatic Segmentation for Music Classification using Competitive Hidden Markov Models
Publication TypeConference Paper
Year of Publication2000
AuthorsBatlle, E., & Cano P.
Abstract abs113 Music information retrieval has become a major topic in the last few years and we can find a wide range of applications that use it. For this reason, audio databases start growing in size as more and more digital audio resources have become available. However, the usefulness of an audio database relies not only on its size but also on its organization and structure. Therefore, much effort must be spent in the labeling process whose complexity grows with database size and diversity.

In this paper we introduce a new audio classification tool and we use its properties to develop an automatic system to segment audio material in a fully unsupervised way. The audio segments obtained with this process are automatically labeled in a way that two segments with similar psychoacoustics properties get the same label. By doing so, the audio signal is automatically segmented into a sequence of abstract acoustic events. This is specially useful to classify huge multimedia databases where a human driven segmentation is not practicable. This automatic classification allow a fast indexing and retrieval of audio fragments. This audio segmentation is done using competitive hidden Markov models as the main classification engine and, thus, no previous classified or hand-labeled data is needed. This powerful classification tool also has a great flexibility and offers the possibility to customize the matching criterion as well as the average segment length according to the application needs.

preprint/postprint documentfiles/publications/ismir2000-eloi.pdf