Compensating changes in speaker position for improved voice-based human-robot communication

Acoustic perturbation due to reverberation and the changes in speaker position are detrimental to seamless human-robot speech-based communication. These cause a mismatch between the speech features at runtime condition and the acoustic model (training condition). Then the degradation of the Automatic Speech Recognition (ASR) and the Spoken Language Understanding (SLU) performances is imminent. As a consequence, the robot fails to understand the spoken commands which will negatively impact interaction experience. In this paper, we propose a framework to improving speech-based human-robot communication in various reverberant environments. The framework is based on robust robot audition that addresses the mismatch problem, striking a balance between technology and the limitations in a real robot setting. Our method improves both ASR and SLU performances. Moreover, the proposed framework has the ability to evolve in minimizing mismatch without human supervision. We experiment with data collected in real environment conditions.