Distant-talking speech recognition with microphone-array sound pickup and NN/MLLR environment equalization

Most contemporary speech recognizers are designed to operate with dose-talking speech and they work best in a quiet, matched training/testing condition. The objective of the paper is to explore utility of existing speech recognition technology in adverse "real-world" environments for distant-talking applications. More specifically, a microphones array (MA) is utilized as effective sound capture to mitigate environmental interference introduced by reverberation and ambient noise, and environment equalizers are employed to approximate a matched training/testing condition for the recognizer such that it need not be retained. Both neural network (NN) and maximum Likelihood Linear regression (MLLR) techniques have been used for environment equalization. Experimental results show that the combined system of MANN-MLLR is able to elevate recognition performance of distant-talking to an extent which is competitive with a retrained speech recognizer.