Robust ego noise suppression of a robot

This paper describes an architecture that can enhance a robot with the capability of performing automatic speech recognition even while the robot is moving. The system consists of three blocks: (1) a multi-channel noise reduction block comprising consequent stages of microphone-array-based sound localization, geometric source separation and post filtering, (2) a single-channel template subtraction block and (3) a speech recognition block. In this work, we specifically investigate a missing feature theory based automatic speech recognition (MFT-ASR) approach in block (3), that makes use of spectrotemporal elements that are derived from (1) and (2) to measure the reliability of the audio features and to generate masks that filter unreliable speech features. We evaluate the proposed technique on a robot using word error rates. Furthermore, we present a detailed analysis of recognition accuracy to determine optimal parameters. Proposed MFT-ASR implementation attains significantly higher recognition performance compared to the performances of both single and multi-channel noise reduction methods.