TitleAutomated Music Analysis Using Dynamic Graphical Models
Publication TypeThesis
Year of Publication2005
AuthorsVogel, BK
AdvisorJordan, M
Academic DepartmentEECS
Number of Pages131
UniversityU.C. Berkeley
CityBerkeley, CA

Music transcription refers to the process of listening to a piece of music and generat- ing the underlying musical score. This is a task that has traditionally required a musician trained in transcription. A robust automated music transcription system would have appli- cations to the music information retrieval (MIR) community, such as automated indexing of musical recordings and query by humming, for example. Good results on the automated music transcription problem might also carry over to the related problem of automated speech recognition. An automated transcription system would also be a useful component in software that teaches one how to play an instrument by allowing the software to tran- scribe performances and then provide feedback. Other applications include transcription of improvised performances and automated score following.The problem of polyphonic transcription with instrument identification is a topic that has not been explored in the literature. In this thesis, we propose and implement dynamic graphical models (DGMs) for multi-instrument polyphonic transcription. The graphical models formalism provides for families of probability distributions to be modeled in a con- cise, modular, and intuitive manner. This thesis is concerned with making some progress on the problem of automated polyphonic multi-instrument music transcription. By multi- instrument transcription, we mean a system capable of listening to a recording in which two or more instruments are playing, and identifying the notes played by each instrument. Our transcription system models both the musical signal behavior as well as some musi- cal structure. We use an instrument-specific timbre model for musical signals that models both the spectral behavior and the variation in the overall sound intensity with time within a note event. We also illustrate the modularity of the graphical models approach by mak- ing some modifications to our multi-instrument transcription system to create a system for guitar transcription.