Yukiko Nakagawa and Hiroshi G. Okuno, Japan Science and Technology Corporation; Hiroaki Kitano, Japan Science and Technology Corporation and Sony Computer Science
We present a method of improving sound source separation using vision. The sound source separation is an essential function to accomplish auditory scene understanding by separating stream of sounds generated from multiple sound sources. By separating a stream of sounds, recognition process, such as speech recognition, can simply work on a single stream, not mixed sound of several speakers. The performance is known to be improved by using stereo/binaural microphone and microphone array which provides spatial information for separation. However, these methods still have more than 20 degree of positional ambiguities. In this paper, we further added visual information to provide more specific and accurate position information. As a result, separation capability was drastically improved. In addition, we found that the use of approximate direction information drastically improve object tracking accuracy of a simple vision system, which in turn improves performance of the auditory system. We claim that the integration of vision and auditory inputs improves performance of tasks in each perception, such as sound source separation and object tracking, by bootstrapping.