Parameter Tuning of Robot Audition Using Speech Recognition System as Evaluation Function

This paper describes parameter tuning of robot audition combining spectral subtraction and speech recognition system. Noise reduction is a key factor for robot audition. It requires some tuning depending on the environment, and we normally set the parameter of noise reduction system so that the error between the objective signal and the signal after noise reduction becomes minimal. However, the objective signal is usually unknown in practice, and the recognition rate is not always the best when the error is minimal. In this paper, spectral subtraction is selected as a simple noise reduction method to evaluate our approach, and the performance of the proposed method is investigated.