An RNN and CRNN Based Approach to Robust Voice Activity Detection
Guan-Bo Wang, Wei-Qiang Zhang · 2019
In this paper, we propose a voice activity detection (VAD) system, which combines a convolutional recurrent neural network (CRNN) and a recurrent neural network (RNN). In order to improve the performance of our system in low signal-noise ratio conditions, we also add a speech-enhancement module, a one-dimensional dilation-erosion module, and a model ensemble module, all of which contribute significantly. We evaluate our proposed system on development dataset of Public Safety Communications (PSC) and Video Annotation for Speech Technologies (VAST) from NIST Open Speech Analytic Technologies 2019 (OpenSAT19). Compared to the baseline system, our proposed system achieves better performance, using OpenSAT19 official evaluation metrics.