Environment Compensation Using A Posteriori Statistics for Distant Speech-based Human-Robot Interaction

Distant hands-free interaction with robot using voice is a difficult task. As the speech signal is reflected within the physical constraints of the room, the speech is contaminated by the environment and speech enhancement alone is not enough to rectify this mismatch in condition. The use of adaptation techniques may result to a robust model adapted to the new environment condition using some small amount of adaptation data. However, the dynamic nature of the interaction between the human and the robot (e.g., constant change in positions) implies a dynamic change in environment condition. Adaptation data collection corresponding to every change may become impractical. In this paper we propose a compensation mechanism short of being an adaptation technique. Our method can cope with the change in environment condition without the use of actual adaptation data. First, we estimate the a posteriori statistics, known in advance and empirically verified to perform well. Then, we infuse the a posteriori statistics to the problematic statistics (i.e., due to environment change) to give a boost to the latter. Weights are selected offline by minimizing the error of the envelopes between the posteriori and the problematic statistics for efficient estimation and fast update. Experimental results in real human-robot environment conditions show that our proposed method performs better than existing adaptation techniques in most cases without the need of adaptation data.