Signal Processing Techniques Applied to Automatic Music Transcription (Slides of PhD Thesis)

In these three years attending the Doctoral School in Telematics and Information Society at University of Florence, I mainly focused my research activity on the topic which gives the title to this Ph.D. thesis: study, design and validation of signal processing techniques applied to automatic music transcription. Automatic transcription of music (AToM) is a difficult problem, which remains still unresolved in some of its application contexts (such as polyphonic music and multi-instrumental transcription). This task refers to the analysis of a digital acoustic or synthesized musical signal, in order to write down pitch, onset time, duration, intensity and source of each sound that occurs in it. In many other research areas such as computer vision and semantic web indexing, efforts are made to automatize several types of human cognitive processes. Computer vision, for example, deals with identifying techniques and strategies for acquiring, processing and understanding images from the real world, in order to produce symbolic/numeric information or decisions rules, which is a daily trivial operation for most of the people. The distinctive aspect, which makes automatic transcription of music such a challenging task, is that its outcome result is a hardly achievable goal not only for ordinary people at large, but even for expert and well trained musicians. This fact can be partially explained by the high degree of perceptual fusion characterizing the human auditory system, according to which we perceive simultaneous and multitimbral sounds as a single entity. Furthermore, the lack of knowledge on human brain processes underlying this complex activity (though the functioning of inner ear transcoding mechanisms have been pervasively studied and understood), justifies the large variety of methods and approaches proposed, ranging from signal processing techniques to higher-level musicological models.