Restoring Lost Speech Components with Generative Adversarial Networks for Speech Communications in Adverse Conditions

摘要

Speech enhancement has been widely implemented to restore quality of speech in communications between humans or between human and machine. For different speech communication scenarios with specific channel and environmental conditions, the types and degrees of speech distortion could vary significantly and many speech enhancement strategies have been developed accordingly. This study deal with a severe distortion problem, i.e., part of the spectral and/or temporal components of the speech are lost completely. The spectral loss is simulated by a transmission channel with very narrow passing bandwidth (lower than 2 kHz) which results in severely degraded speech quality; the temporal loss is simulated by packet loss up to 50% percent in massive communication which results in poor speech intelligibility. A generative adversarial networks (GAN) based speech enhancement scheme is proposed for restoring the missing spectral and temporal components with different network structure and parameters. A set of experiments have been conducted to evaluate the effectiveness of proposed enhancement scheme and promising results have achieved.