Gökhan Ince, "Noise Reduction During Ego-motion of the Head", 2007.

An active auditory perception system is very essential for robots to be able to interact with their environment. Tasks like sound localization and speech recognition have to be performed with high accuracy even when the head (or whole robot) is moving. However, the movement of the head inevitably generates noise due to its motors. This problem is very crucial, because the motors are located closer to the microphones than the sound sources. Considering that a head rotating with higher velocity causes more noise energy, this undesirable noise has to be suppressed. Noise reduction can be accomplished only if the noise can be predicted instantaneously. If the motions of the head are known a priori, the noise can be estimated since a certain type of motion produces almost the same pattern every time it is performed. Using this information, this master thesis investigates spectral subtraction methods utilizing various algorithms that build on a suitable template for the respective motor noise. Five single-channel enhancement techniques are analyzed and they are optimized with respect to their corresponding parameters. Moreover, their performances are compared with a newly proposed onset measurement criterion as well as the traditional evaluation methods like perceptual quality tests and automatic speech recognizer (ASR). The results of all evaluation methods showed the effectiveness of the proposed idea. The problems of the current approach are the precise synchronization of sound signal and motor noise template and the large number of templates needed to model all possible motions. Therefore, a more flexible method which can overcome the timing problem of the template and reduce the size of the template database containing motion templates in the 3-D space is proposed. The very first steps taken to model the noise for each time instant by matching a pre-recorded motor noise depending on the position/velocity/acceleration information of the motor are explained. Finally, the limitations of the present spectral subtraction methods have been identified and suggestions for further improvement have been put forward.