A Study on Improving Acoustic Model for Robust and Far-Field Speech Recognition

Abstract

Far-field speech recognition is an essential technique for man-machine interactions. It aims to enable smart devices to recognize distant human speech. This technology is applied to many scenarios such as smart home appliances (smart loudspeaker, smart TV) and meeting transcription. Despite the significant advancement made in robust and far-field speech recognition after the introduction of deep neural network based acoustic models, the far-field speech recognition remains a challenging task due to various factors such as background noise, reverberation and even human voice interference. In this paper, we describe several technical advances for improving the performance of large-scale far-field speech recognition, including simulated data generation, improvements on front-end modules and neural network based acoustic models. Experimental results on several Mandarin Chinese speech recognition tasks have demonstrated that the combination of these technical advances can significantly outperform the conventional models.