A Unified Front-end Anti-interference Approach for Robust Automatic Speech Recognition

Yunming Liang, Yi Zhou, Yongbao Ma, Hongqing Liu

Abstract: Front-end technique has become an indispensable part for robust automatic speech recognition (ASR). It was recently reported that Deep Xi (a deep learning approach to a priori SNR estimation) is used as a front-end tool for ASR due to its high speech enhancement performance to significantly improve the robustness of ASR systems. However, Deep Xi is not suitable for processing speech signals contaminated by musical instrument interference which is commonly encountered in daily life. To solve this problem, this paper proposes a new effective method unifying independent low-rank matrix analysis (ILRMA) and Deep Xi to design a front-end anti-interference ASR system in the presence of musical instrument interference. Experimental results show that compared with the conventional Deep Xi, the proposed method has better performance in terms of the robustness of ASR system.