情報処理学会 第73回全国大会講演要旨

This paper describes a speaker localization system using two microphones
embedded in the head of the SIG-2 humanoid robot. Nowadays, studies based on
microphone arrays utilize many microphones to obtain high detection
performance in acoustical engineering. However, using multiple microphones
would raise costs and hinder maintenance. To achieve good results in speaker
localization only with two low-cost microphones, we employ the statistical
model-based voice activity detection (VAD) and a sound-source-direction
estimator based on the generalized cross-correlation with phase transform
(GCC-PHAT). The experimental results show that SIG-2 can localize a caller
reliably when the speaker calls robot's name in the azimuth.