Teigo Nakamura

Analysis of Composite Corridors

This work began as an attempt to find and catalog the mean values and temperatures of a well-defined set of relatively simple common Go positions, extending a similar but smaller catalog in Table E.10, Appendix E of the book Mathematical Go [1].

Continuous inaudible recognition of Japanese vowels using features detected at lips shape peaks based on surface electromyography

We are working the recognition of inaudible Japanese speech by using surface electromyography. During natural speech, muscle activity changes continuously and gently under the influence of the preceding or succeeding phoneme, even within the utterance of a single vowel. We need to detect the temporal positions of such changes to extract stable features of phonemes. In our previous works, we expected that time position, which is the most distinct feature of a vowel, is the time when the lips are formed clearly. We used CoB, which refers to the center of gravity of wavelet coefficients, to evaluate muscular activity for tracking lips shape. However, the method was complex in terms of execution. It assumes that CoB is a insufficient information for the evaluating muscular activity. CoB is inferior in correlation with muscle strength, because it has frequency features which reflect nervous tissue activity. Therefore, we evaluate muscular activity by using a frequency feature and a signal power feature, and track lip shape. In this paper, we describe a method for tracking lip shape and detecting lip shape peaks, and report the recognition result of the continuous utterance of a pair of inaudible Japanese vowels by using the features around such peaks.

Document Classification with Varied Viewpoints Using Matrix Decomposition

The result of document classification is not unique and there can be many other results from the different perspectives. In other words, the classification results can vary according to users viewpoints of classification. If a document classification system ignores the users viewpoints, the result of classification will be different from the result desired by the user and the difference between the users desired result and the systems result can cause some inhibitions and oversights in information retrieval. So, we extract the users viewpoints from the classification examples performed preliminarily by the user and use them to the following classification in order to reflect the users desire. In this paper, we propose four methods to extract viewpoints and three methods to classify documents using matrix decomposition such as Nonnegative Matrix Factorization (NMF). We show the results of comparative experiments with the original NMF, Semi-Supervised NMF (SSNMF) and our proposed methods.