【摘要】 随着语音通信和人机语音交互系统的不断普及,人们越来越期待抛开话筒和耳机等繁琐的设备,实现像人类对话一般自然的人机语音交流.然而,语音毕竟只是一种声波,在空气中传输时难免受到各种影响,例如声波的衰减、墙壁和障碍物的多次反射以及同时存在的其他声源等.如果不采用近讲的拾音方式,那么这些因素都会对传播中的语音声波造成干扰.特别是当多个语音系统和多个说话人处于同一环境时,如何确保系统正确接收语音信息,决定了语音系统能否走向实用.本文参考人类的听觉注意机理,提出充分利用对目标语音及干扰声源的先验知识,检测和提升目标语音,并通过将传声器阵列、语音唤醒、目标语音检测、语音增强、混响抑制等一系列技术相结合,实现抗干扰的目标语音拾取.更多还原

【Abstract】 Currently, a natural speech-picking mode is badly needed in speech communication and in humancomputer interaction systems. However, speech is usually corrupted by attenuation, multi-path propagation, and various interferences before it is received, especially when there exist several speech systems and users. It is important for practical speech systems to pick the correct speech signal within complex environments. In this paper, the mechanism of auditory attention ability is simulated through a target speech-picking system in which the a priori knowledge of the target speech and interference of sound sources are used carefully to detect and improve the target speech. The technologies of microphone arrays, wake-up-words, target speech detection, speech enhancement, and dereverberation are combined in this strategy to fulfill the task of robust target speech-picking.更多还原