Automatische Erkennung von Liedabschnitten in Musikstücken

Diploma thesis (pdf 5.971 KB)

A new class of algorithms tries to find similarities between audio-files which a human listener would intuitively recognize. These algorithms are becoming more and more important as they offer new and innovative possibilities of searching through music collections, generating (automatic) playlists, or visualizing similarities between audio files.

In the diploma thesis, a system that is capable of analyzing audio-data based on both signal-theoretical and music-theoretical principles will be developed. After a preprocessing and feature extraction step, the system will classify the retrieved information to find similarities between audio-files which a human listener would intuitively recognize. The integration of music-theoretical knowledge allows a segmentation of audio-files into musically relevant sections (like beat/measure, song segments). This segmentation allows for a section-aligned analysis which is assumed to improve the performance of the basic feature extraction algorithms and should thereby enhance the ensuing classification step.

The planned system will mainly consist of the following steps:

• preprocessing
• extraction of spectral information and detection of harmonic progression
• automatic song segmentation
• extraction of features within certain song segments
• classification of extracted features
• visualization of classified information