[様式－学 5]
Abstract of Doctoral Thesis
Title ： Real-time Sound Source Localization in Real Environments
Doctoral Program In Information Science and Engineering Graduate School of Information Science and Engineering Ritsumeikan University
ふりがな（ひらがな表記）はやしだ こうへい 氏 名（姓・名 アルファベット表記）Hayashida Kohei
Sound source location that is estimated from acoustic signals is useful information for various applications such as the automatic abnormity detection in blind area of security camera, the speech rejection from non-user at distant position for speech interface, and so on. Required location information is different by the application. In automatic video camera control for abnormal sound, accurately sound source direction and distance are necessary. In that case, the sound source localization methods using plural microphones are suitable. On the other hand, accurately sound source direction and distance are unnecessary for desired/ undesired talker discrimination in speech interface. Whether the talker would be within the certain distance is sufficient location information for this purpose. Close/distant talker discrimination method with single microphone suits these cases better than sound source localization with plural microphones. However, real-time processing is difficult in both conventional sound source localization methods with single and plural
microphones.
In this thesis, the author proposed (I) multi-resolution scanning in spatial and frequency domains for sound source localization method with plural microphones, and (II) close/distant talker
discrimination method with single microphone based on kurtosis of linear prediction residual signals for real-time processing. As component (I), conventional sound source localization methods with plural microphones are difficult to localize a sound source in real time with higher spatial resolution.
To overcome this problem, we proposed a new localization method with different spatial resolution in spatial and frequency domains. Finally, as component (II), conventional close/distant talker discrimination method with single microphone is difficult to realize real-time processing because this method needs higher-order linear prediction analysis. In this thesis, we proposed a new method to discriminate close-talking speech from distant-talking speech with a single microphone based on the kurtosis of the linear prediction residual signals, and it can be calculated with lower-order linear prediction analysis. The experimental results in real environments revealed that the both proposed methods could localize sound source accurately in real time.
