数字内容技术与服务研究中心知识库

In this paper, we describe our ZTSpeech for two tracks of CHiME-5 challenge. For front-end, our experiments conduct the comparisons between several popular beamforming methods. Besides, we also propose a omnidirectional minimum variance distortionless response (OMVDR) followed by weighted prediction error (WPE). Furthermore, we investigate the impact of data augmentation and data combinations.  For backend, several acoustic models (AMs) with different architectures are deeply investigated. N-gram-based and recurrent neural network (RNN)-based language models (LMs) are both evaluated. For single-array track, by combining the most effective approaches, our final system can achieve 11.94% promotion on performance in evaluation set, from 73.27% to 61.33%. For multiple-array track, our final system can achieve 12.29% improvement in evaluation set, from 73.30% to 61.01%.