Evaluation of human-phonatory radiation characteristics with a polyhedron loudspeaker

Spoken dialog systems have been studied for car navigation systems and voice search systems. For evaluating these systems, a loudspeaker is used instead of a human because these systems require various kinds of speech samples. However, the sounds radiated by loudspeaker cannot reproduce human-phonatory radiation characteristics. Therefore, the mouth simulator is utilized to reproduce human-phonatory radiation characteristics. Although it is based on the average mouth shapes, shapes of mouth are different among phonemes. Therefore, due to the hardware structure, it cannot accurately reproduce various human-phonatory radiation characteristics affected by shapes of mouth. In this study, we developed a polyhedron loudspeaker to solve this problem. It consists of 11 loudspeakers, which are independently controlled. Controlling eleven loudspeakers makes it possible to reproduce desired radiation characteristics. By utilizing this method, we try to reproduce human-phonatory radiation characteristics of Japanese five vowels and typical consonants with digital filters which were adaptively designed. We carried out an evaluation experiment in various measuring points to verify the effectiveness of the proposed method. As a result, it was confirmed that human-phonatory radiation characteristics with the proposed method could be accurately approximated compared with the conventional mouth simulator.