[Back]

G. Marzorati:
"An Investigation of Piano Transcription Algorithm for Jazz Music";
Supervisor: F. Antonacci, P. Knees, R. Vogl; Politecnico di Milano, Italy, 2018.

The thesis aims to create an annotated musical dataset and to propose an Automatic Music Transcription system specific to jazz music only. Although many available annotated datasets are built from the audio recordings, the proposed one is built from MIDI file format data, providing robust annotation. The automatic polyphonic transcription method uses a Convolutional Neural Network for the prediction of the outcome.
Automatic Music Transcription is an interesting and active research field of Music Information Retrieval. Automatic Music Transcription refers to the analysis of the musical signal to extract a parametric representation of it, e.g. a musical score or MIDI format file. Even for man, the transcription of music is difficult and still remains a hard task requiring a deep knowledge of music and high level of musical training. Providing a parametric representation of audio signals would be important for application to annotated music for automatic research in large and interactive musical systems. Massive sup- port would be given to the musicology fields producing annotation for audio performance without any written representation, and to the education field. The work hereby presented is focused on the jazz genre, due to its variety of styles and improvisation parts, of which usually there is no available transcription, and to which the field of Automatic Music Transcription can be of help. Its variability makes the problem of Automatic Music Transcription even more challenging and also for that reason there is not much work available.
Results of the transcription system highlighted the difficulties of transcribing jazz music, compared to classical music, but still comparable to state-of-art methodologies, producing an f-measure of 0.837 testing the Neural Network on 30 tracks of MAPS dataset and 0.50 from the jazz dataset experiment.