The Computer Music Project at Carnegie Mellon is collaborating with the MusEn Project at the University of Michigan on a large-scale project to investigate content-based music information retrieval. The joint system we built is called MUSART, for "Music Analysis and Retrieval Technology." Our group has studied broad issues of audio analysis, the development of new representations, especially those dealing with musical style, and the engineering of large-scale musical databases.

Some of the goals of this research project are:

• Construct a content-based retrieval system for digital-music recordings and an explication of the engineering principles behind it.
• Investigate representations of style.
• Characterize music representations from the standpoint of the loss of information.
• Develop user interfaces for music searching, which allow query resynthesis and evaluation of results without musical training. Develop mechanisms for audio skimming.

In many ways, this work represents a "grand challenge" for computer music: we want to accept acoustic input, perform a variety of analyses, including polyphonic transcription, harmonic analysis, rhythmic analysis, and textural analysis. Once music is in symbolic form, we want to explore further analysis to identify motives, characteristics common to a composer or genre, and generally parse the music into a variety of structured representations. In order to understand these representations, we must be able to reverse the process, that is, compose and synthesize music. Thus, the work encompasses the "end-to-end" problem from analysis to high-level-representations and back to synthesis.

This project was funded by the National Science Foundation from 2000 to 2003. The work begun in this project continues to influence and inspire much of what we do.