15 Nov 2016

Dr. Eita Nakamura (Kyoto University, Japan) and Dr. Shinji Sako (Nagoya Institute of Technology, Japan)
will be giving two talks:

"Rhythm Transcription of Piano Performances Based on Hierarchical. Bayesian Modelling of Repetition and Modification of Musical Note Patterns" by Dr. Eita Nakamura. Kyoto University, Japan. (15h Nov, 17:00h. Room 52.321)

Abstract
We present a method of rhythm transcription (i.e., automatic recognition of note values in music performance signals) based on a Bayesian music language model that describes the repetitive structure of musical notes. Conventionally, music language models for music transcription are trained with a dataset of musical pieces. Because typical musical pieces have repetitions consisting of a limited number of note patterns, better models fitting individual pieces could be obtained by inducing compact grammars. The main challenges are inducing appropriate grammar for a score that is observed indirectly through a performance and capturing incomplete repetitions, which can be represented as repetitions with modifications. We propose a hierarchical Bayesian model in which the generation of a language model is described with a Dirichlet process and the production of musical notes is described with a hierarchical hidden Markov model (HMM) that incorporates the process of modifying note patterns. We derive an efficient algorithm based on Gibbs sampling for simultaneously inferring from a performance signal the score and the individual language model behind it. Evaluations showed that the proposed model outperformed previously studied HMM-based models.
 

"Real-time audio-to-score following and its applications" by Dr. Shinji Sako (and his students). Nagoya Institute of Technology, Japan. (15th Nov, 17:45 h. Room 52.321)

Abstract
We present a robust on-line algorithm for real-time audio-to-score following based on a delayed decision and anticipation framework. We employ Segmental Conditional Random Fields and Linear Dynamical System to model musical performance by human. The combination of these models allows an efficient iterative decoding of score position and tempo. The combined advantages of our approach are the delayed-decision
Viterbi algorithm which utilizes future information to determine past score position with high reliability, thus improving alignment accuracy, and the fact that the future position can be anticipated using an adaptively estimated tempo. We also talk about interim progress of the research and some applications by using this
technique.