| Research | Dissertation | Other Publications |
When presented with musical sounds, humans take advantage of prior knowledge of acoustic and musical context to accomplish an impressive array of cognitive listening tasks, such as meter tracking, transcription, style classification, instrument identification, harmonic analysis, and melody prediction. This dissertation presents an extensible Bayesian inference engine capable of performing such tasks. Because the framework facilitates the modeling of listeners with differing prior expectations, the predictions and decisions made by different listeners can easily be compared.
Although a listener with specific experience could simply be created by learning prior distributions directly from a given corpus, a more interesting approach is to construct listeners whose expectations are governed by rules of music theory. Such rules are often expressed as statements involving musical tendencies, e.g., "A large upward melodic interval is typically followed by a smaller downward interval." This dissertation focuses on a novel method of automatically transforming rule sets into maximum entropy stochastic grammars suitable for use in dynamic Bayesian networks. Encoding rule-based expectations allows the system to infer which rules are most responsible for predicting musical attributes at each time in a piece, and to identify which rules are violated at points of musical surprise.
While a variety of interesting tasks can be performed using symbolic data as input, the framework also includes a layer allowing inference to be done using recorded audio signals as input. In the signal processing layer, acoustic expectations are encoded by modeling the spectrotemporal evolution of instrument tones, and the signal is segmented into a sequence of note events. A system in which signal and symbolic layers inform one another is desirable because musical expectations can help the system compensate for corrupted signals, and the ability to predict musical sequences suggests a future Markov Chain Monte Carlo (MCMC) implementation in which the sampling distribution concentrates on the most likely transitions, thereby avoiding the computational cost of evaluating all points in the potentially vast space of possible transitions.
Download: PDF
| Research | Dissertation | Other Publications |