Automatic Segmentation and Classification of Articulation in Monophonic Music

Automatic music transcription has attracted a great interest from computer scientists and musicians for more than thirty-five years. The process of transcription includes extraction of fundamental frequencies, and segmentation of the continuous audio signal by detecting onsets and offsets of sound events. In a following step, quantization of pitch and timing gives a symbolic representation of the performance, often presented as a score or MIDI file. At best, the musician’s interpretation as manifested in variations in tempo (accelerando-ritardando), phrasing, and articulation (staccato, legato) may be captured as well. Automatic transcription can be used for documenting recordings, comparing a performance with the original score and music retrieval when searching for music on the Internet. In this thesis, three programs for automatic transcription are first evaluated and compared in order to estimate the accuracy, in particular with regard to onset detection. In a second part, articulation in recorder playing is studied by timing analysis of recorded student and teacher performances combined with listening tests.