Project Description      Project Parts      Masters Projects      Dissertations     Publications     Queries

Project Description

The first part of the MiDiLiB-project aims at designing and realizing a large digital library of classical music. Compared to conventional music libraries this digital library is going to offer a number of new features: world-wide access, novel content-based retrieval and browsing techniques, multimedia processing of documents with suitable tools for analysis, editing and visualization with a view to teaching, research and presentation. To reach these goals, a number of problems both in research and in development have to be solved. Among these are the construction of an overall system including a web-based user interface as well as the development of a MIDI database. One of the main research problems is to discover (semi-)automatic procedures for content-based indexing and retrieval of melodies in classical music pieces, allowing an efficient content-based search. Here, we are faced with a number of problems that are presently not solved in a satisfactory way: which parts of a musical piece should be put into an index, what are the musical keywords (themes, motifes), what are suitable data structures for a content-based index, how can be searched efficiently and in a fault-tolerant way in such an index, how can a user's search be interactively supported?

By this time we have implemented a prototypical platform with a web-based user interface. Our MIDI database contains over 20,000 MIDI files. As most of these files result from playing the actual piece on a MIDI-piano, we are faced with the problem of normalization and quantization of MIDI files. In this context we have developed encoding schemes that support the processing of search queries. Furthermore, we have designed a parametric quantization model that allows to adjust the degree of fault tolerance to the specific application in mind. Another fundamental problem is that of melodic similarity. We have studied several melodic distance measures. In simple situations we have applied such measures to analyze fugues and melodic variations. With respect to fault tolerant retrieval the algorithms have to be both time and space efficient. Here we have developed a new encoding of melodies that simultaneously reflects pitch and rhythm. String-based melodic similarity evaluation is then closely related to the longest common subsequence problem for which we have designed a new practical and very efficient algorithm.

The extraction of melodies from polyphonic pieces is another difficult task. We have developed variants of skyline-like algorithms to obtain main melodies of a musical piece. This approach is currently improved by taking global aspects into account (Gestaltpsychology, musicology).