Performance evaluation of meaning-based parameter tuning of robot audition

In this paper, meaning-based parameter tuning of robot audition is investigated. Noise reduction system often requires parameter tuning depending on the environment. The typical parameter tuning is executed by error minimization between the target signal and the filter output after noise reduction. However, the target signal is generally unforeknown in advance, and the minimized error does not assure the high recognition rate of speech recognition system. We employ speech recognition system as an evaluation function of parameter tuning to solve the problems. We implemented the system to a human-size humanoid and verified the effectiveness of the proposed method on an experimental basis.