FWF Translational Research L511-N15 - Department of Computational Perception

Department of

Computational Perception

Home – Mission – Teaching – People – Research – Media – Impressum

MUSIC RETRIEVAL BEYOND SIMPLE AUDIO SIMILARITY

Project Title: Music Retrieval Beyond Simple Audio Similarity

Sponsor: Austrian National Science Fund (Fonds zur Förderung der wissenschaftlichen Forschung, FWF)

Project Number: L511-N15

Duration: 48 months (May 2008 – April 2012)

In cooperation with: Austrian Research Institute for Artificial Intelligence (OFAI), Vienna – Intelligent Music Processing and Machine Learning Group

Persons involved:

Gerhard Widmer (Project Leader)
Peter Knees
Tim Pohle
Klaus Seyerlehner
Dominik Schnitzer
Reinhard Sonnleitner
Florian Krebs
Markus Schedl

Abstract

Digital music archives have reached capacities nowadays that require 'intelligent' computational methods to assist the user in finding and retrieving desired music. Responding to these demands, the still growing research field of Music Information Retrieval (MIR) develops techniques to facilitate access to music. The previous FWF project L112 "Operational Models of Music Similarity for Music Information Retrieval" enabled us to deepen our expertise in intelligent music audio processing and assessment of musical similarity, and to develop computational methods that are now finding their way into practical applications. Additionally, we undertook first steps towards extracting relevant (meta-)information on the work of musical artists from the Internet. The goal of this new project is to substantially advance this latter work on Web-based music information retrieval, and to develop robust methods that can be used as a basis for commercially relevant application projects.

In particular, we will carry out research to accomplish 4 main goals:

to overcome the dependency on commercial Web search engines that are currently necessary to obtain musically relevant Web pages;
to improve existing Web-based MIR techniques to acquire contextual information and to make them more robust by learning to deal with ambiguous information;
to develop methods that permit representation of music pieces and artists in a "semantic" space that is characterized by meaningful terms and descriptions;
to automatically discover semantic relations beyond pure acoustic similarity in order to extend retrieval capabilities and improve current similarity measures.

The outcome of this research will be a set of Web-based methods to derive manifold types of music-related information, independent of commercial search engines. Using these methods, many new services will become possible, such as music information systems that autonomously collect and edit artist related data and discover connections between them, or music search engines that can be queried through meaningful natural language expressions. Existing content-based music services will also benefit from these complementary sources of information.

Publications

Indexing Content-Based Music Similarity Models for Fast Retrieval in Massive Databases

D. Schnitzer

Dissertation, Johannes Kepler University Linz, October 2011.