Mark Levy

Contact Details

Title: Research Assistant
Tel: Internal: [13] 5528
National: 020 7882 5528
International: +44 20 7882 5528
Fax: National: 020 7882 7997
International: +44 20 7882 7997
Email: mark.levy [at] elec.qmul.ac.uk
Room: Eng. 112

Research Group: Centre for Digital Music

Research Project: Online Music Recognition and Search 2 (EPSRC)

I build prototype search engines for music. I try to learn a Latent Semantic Space for music by looking at millions of descriptions - supplied by listeners as social tags - for many thousands of tracks. This means learning how to represent each track as a point in a multi- dimensional space (typically with around 100 dimensions), where each axis represents some musical concept, and the position of a track along an axis shows how relevant this concept is to the track in question. The concepts or Aspects aren't defined in advance as genres, styles, moods, etc.: instead we assume that they are latent (lying hidden) in the data, and we learn them by discovering which words tend to be applied to the same tracks. The latent space forms the basis of powerful search engines because tracks whose points lie close together have meaningful characteristics in common, so we can find tracks similar to a given one, as well as looking up tracks that are a good fit for a conventional query like "soft sexy jazz". I'm currently extending my model to include audio information as well. This means that it tries to learn the characteristic sound of each latent aspect, which helps with tracks for which we don't have very full descriptions.

Previous Research Project: Hierarchical Segmentation and Semantic Markup of Musical Signals (EPSRC)

I work on automatic segmentation of musical audio i.e. finding and labelling sections within a track directly from the audio data. I am working both on segmentation methods, and developing prototype applications for consumers and for professionals such as recording engineers. These include applications to search large music collections for tracks similar to a supplied query, to extract representative audio 'thumbnails' (typically the chorus of a song) from individual tracks or collections, and to navigate automatically from one section of a piece to another within an audio editor. Most previous work on segmentation has been based on finding points of change in sequences of features extracted from the audio, then deciding which ones are genuinely segment boundaries, and finally attempting to label segments which appear to be of a similar type (all verses, instrumental breaks, etc.). Our approach in contrast is model-based. We express our assumptions about the sort of structure we expect to find in a music track in the form of a model parameterised by suitable probability distributions, for example describing the expected length of the sections. The model can then be trained on the audio features, and the structure inferred from the trained model, typically as the most likely sequence of model states to have generated the observed features.

Before arriving at QMUL in 2005 I worked as a software developer and as a professional musician. I continue to give concerts and make recordings as time allows and you can find more information about my musical activities here.

Publications

This material is presented to ensure timely dissemination of scholarly and technical work. Copyright and all rights therein are retained by authors or by other copyright holders. All persons copying this information are expected to adhere to the terms and constraints invoked by each author's copyright. In most cases, these works may not be reposted without the explicit permission of the copyright holder.

The following copyright applies to any articles on this page published by IEEE: "©20xx IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works must be obtained from the IEEE."

G. Kreutz, M. Levy and M. Sandler, Emotion words in social tags for popular music, submitted to Music Perception 2008

M. Levy and M. Sandler, A semantic space for music derived from social tags, submitted to ISMIR 2007

M. Levy and M. Sandler, Structural segmentation of musical audio by constrained clustering, submitted to IEEE Trans. ASLP

K. Noland, M. Levy and M. Sandler, A comparison of timbral and harmonic music segmentation algorithms, In Proc, ICASSP 2007

M. B. Sandler and M. Levy, Signal-based music searching and browsing, Proc. ICCE 2007

M. Levy and M. Sandler, Lightweight measures for timbral similarity of musical audio, Proc. ACM Multimedia 2006

M. Levy, M. Sandler and M. Casey. Extraction of High-Level Musical Structure from Audio Data and its Application to Thumbnail Generation. In Proc. ICASSP, 2006.

M. Levy and M. Sandler. Application of Segmentation and Thumbnailing to Music Browsing and Searching. To appear as AES 120th Convention Paper, 2006.

M. Levy and M. Sandler. New Methods in Structural Segmentation of Musical Audio. Submitted to Eusipco, 2006.

M. Levy, M. Sandler and M. Casey. Extracting Musical Audio Thumbnails. Poster for DMRN Workshop, December 2005.