Automatic Segmentation and Labelling

Diploma Thesis (pdf 1563KB)

Information overload from a continuously rising number of sources is becoming increasingly unmanageable, this is also true for music collections. An overview with traditional means is impossible. The answer could be automatic structuring to be able to search specifically and to enable quick information retrieval.

The problem, as opposed to musical data in symbolic representation (MIDI), lies in the need for a pre-processing stage to find segments meaningful for the structure of the piece and finally label them.

This thesis deals with the analysis of the structure of musical data from the genre pop/rock. The first approach uses harmonic analysis to obtain the musical form of the piece, secondly a search for repeating elements is conducted to determine the position of the chorus in the song.

The implementation of modular programming-routines, as well as evaluation, takes place in Matlab software. For harmonic analysis a new model is presented, which combines two previously known approaches. Beat-synchronous calculation of the Chromagram is followed by a Harmonic Change Detection Function, to find segments suitable for the analysis of a musical structure. The second model searches for similarities within the song. The chorus is assumingly the part occurring quite similarly at different positions of a song. With the help of distance matrices and methods borrowed from image-processing and pattern-search applications, the localisation is made possible. Evaluation is performed by comparing manual annotations of musical form with automatic detections.

It will be shown that the presented harmonic model is potentially capable of finding song structures, though problems arise to be solved in future versions of the algorithm. The search for similarities functions well, as long as a basic musical form (e.g. Verse-Chorus-Chorus) is met in the song.