Combining Log-Spectral Domain Compensation with MVA Feature Post-Processing for Robust Speech Recognition

Abstract

In this paper, we present a new scheme combining environment compensation with feature postprocessing to improve the robustness of speech recognition systems. The environment compensation is implemented in the log-spectral domain and the environment model is approximated by Statistical Linear Approximation (SLA). The MVA feature postprocessing is used to deal with the residual mismatch between compensated noisy speech and clean speech. We have evaluated recognition performance under noisy environments using NOISEX-92 database and recorded speech signals in continuous speech recognition task. Experimental results show that our approach exhibits considerable improvements in the degraded environment.