Towards robust telephony speech recognition in office and automobile environments

Abstract

This study is concerned with improving the robustness of our telephony speech recognition system. Our previous implementation of this system handled both landline and cellular speech produced in a relatively quiet environment, such as in a regular office. However, it was found to be unduly vulnerable to background noise. In particular, we wanted to improve the accuracy of the system in the environment of a moving automobile, without impacting its performance in a quieter background. We collected samples of automobile noise under various operating conditions. We used these noise files to artificially generate noisy training data. We applied MAP adaptation procedure to study the value of such noisy training data and found that it helped to improve the robustness of our telephony speech recognition system. In a related study, we also investigated performance scores as a function of the total number of vocabulary entries in the system, as this has implications for a practical system implementation.