Model-based segmentation of time-frequency images for musical transcription.

The musical transcription problem seeks a low-order parametric representation of an audio signal where it is assumed that this signal was generated by one or more musical instruments playing simultaneously. Specialized hardware and software systems exist for transcription of specific instruments playing in isolation, but the general multi-instrument transcription problem has not yet been successfully addressed. It has been observed that single instruments playing notes in isolation leads to highly structured time-frequency descriptions. The desire to exploit this structure has driven the heuristics used in current transcription research. Such structure breaks down, however, as more instruments are added to the signal or it becomes corrupted by reverberation or vibrato, for example. This structural breakdown appears responsible for the poor performance of transcription systems designed for the single-instrument case. In this research, we do away with many of the heuristics of transcription systems, replacing them with statistical signal processing and search algorithms that fundamentally rely upon models of musical instrument behavior. We construct these models based upon statistical characterizations of instrument recordings, which stands in contrast to existing approaches that rely upon only intuitive notions of what features comprise a musical note. We provide an open architecture for optimizing sub-components of the framework and for extending the transcription system to other types of musical sources. Finally, we propose several metrics and test cases for evaluating the performance of any transcription system. We show that our transcription algorithm performs well for both single- and multi-instrument cases.Ph.D.Applied SciencesCommunication and the ArtsElectrical engineeringMusicUniversity of Michigan, Horace H. Rackham School of Graduate Studieshttp://deepblue.lib.umich.edu/bitstream/2027.42/132009/2/9938543.pd