[Japanese | Thesis | Researches in Minoh Lab | Minoh Lab]
In a lecture room students divide into some groups and discuss. We call this 'groupwork' and we propose a method of sound source localization for recording group work.
Since sound source localization is for shooting talkers by automation camera and getting speech, we should only locate the talkers' seat. We make the position measurement problem result in a search problem. A lot of sound source exist in the lecture room. We extract the students' seat by using image from a fisheye camera and narrow the search space. This can enable us to locate multiple sound source. In a general discussion talkers often change and we cannot apply the method using correlation between signals to our case. We use sound pressure.
Ideally every speech has the same quality and we set microphones to the ceiling. If sound source changes, the distance to a microphone and the pressure observed by microphone does not change largely. We use directional microphones.
We consider the combination of seat where potential sound source exists and calculate the pressure generated from these seats based on the model we make. The combination that provides the nearest pressure to actual one is regarded as the location of sound source.
We apply our method to simulated groupworks and examine the validation of our method.