2.2 Robot audition based on sound environment understanding

We have been studying sound environment understanding (Computational Auditory Scene Analysis) [9], focusing on the importance of dealing with general sounds including music, environmental sounds and their mixture with speech. An important task in a study on sound environment understanding is sound mixture processing. Sound environment understanding is not to avoid the problem of sound mixture with a close-talking microphone set on the lips of a speaker, but to tackle the processing of a sound mixture with mixed sound as an input. The three major problems in sound environment understanding are sound source localization, sound source separation and automatic speech recognition in doa recognition. Various technologies have been researched and developed for each of those problems so far. However, all those technologies require some specific conditions to draw their maximum performance. In order to draw the maximum performance in combining these technologies for robot audition, it is essential to systematize interfaces of the individual techniques. In order to achieve this goal, the middleware that can effectively provide a combination with a good balance is also important. The robot audition software HARK is constructed on the middleware called FlowDesigner and provides a function of sound environment understanding on the premise of using eight microphones. HARK is designed based on the principle of removing the need for prior knowledge as much as possible, and is a system that aims at being the “OpenCV of acoustic processing”. Indeed, a robot that recognizes orders for dishes from three different persons and that judge orally-played rock, paper, scissors have been realized. Although images and video pictures are generally the environmental sensors, they do not accept appearance and disappearance and dark places. Therefore they are not always useful. It is necessary to remove ambiguity of images and video pictures by sound information and adversely to remove ambiguity of acoustic information by image information. For example, it is extremely difficult to judge if the sound source is in front or back by sound source localization with two microphones.

\includegraphics[width=0.25\columnwidth ]{fig/Intro/SIG-II.eps} \includegraphics[width=0.3\columnwidth ]{fig/Intro/Kim-Binaural-Nod.eps}
Figure 2.2: Active audition of SIG2: For the surrounding sound, ambiguity of front and back is removed by moving the neck right, left and down.