Robust Sound Source Identification for a Humanoid Robot

Authors: Alexej Swerdlow, Timo Machmer, Benjamin Kühn, Kristian Kroschel

Abstract:

In this paper, investigations regarding the robust classification of acoustically observable sources (kitchen appliances and speakers) are presented. Thereby, two deciding factors are considered. On the one hand, the data acquisition should be exclusively done by the on-board sensors of the robot. On the other side, the entire information processing must be handled in real time due to the requirements given by the robot hardware. In so doing, there are positive and negative characteristics to think about. Since audio data are picked up in a small room like a kitchen with many sound sources, the background noise and reverberation of the signals of interest have to be taken into account. The redundancy of the signals picked up by the microphone array, which is installed on the robot head, can be employed to improvethe classification accuracy. The presented system is based on Gaussian Mixture Models in correspondence with the Mel Frequency Cepstral Coefficients as acoustic signal features. Furthermore, a Universal Background Modelis used for the special case of speaker identification. In this work, studies regarding the channel combination, the necessary length of the training phase, and the minimum data length for the evaluation are presented.