互联网时代语音识别基本问题

语音识别技术经过半个世纪的积累，于近年来达到大规模商用水平．本文概括了统计语音识别理论的发展状况，并单独介绍了深度神经网络在声学建模、语言建模、多语言共享、语义识别等方面的卓越性能．深度神经网络的性能优势引起了我们强烈的兴趣．通过回顾类人听觉信息处理对深度神经网络的改进作用，我们意识到，深度神经网络与类人听觉信息处理相结合，必将推进语音识别技术的进一步发展．反过来，深度神经网络技术在语音识别中的进步，也必将推动类人听觉信总、处理技术的进步．语音识别技术后续发展的重点是对深度神经网络的结构和训练算法的改进使之更好地实现类人听觉．最后，我们分析了采用深度神经网络模拟人类听觉的抗噪修复机理和听觉关注机理的可能性．

Some basic problems of speech recognition in the internet era

Speech recognition technology has reached its highly mature period after half a century of development. In this paper, we first summarize the development of statistical theory and methods of speech recognition, and then introduce the excellent performance of deep neural networks in the tasks of acoustic modeling, language modeling, multilingual acoustic sharing and semantic classification. Deep neural networks＇ performance advances arouse our great interest. By reviewing the pushing effects on deep neural networks of human-like auditory in- formation processing, we realized that the combination of them will promote the further development of speech recognition. In turn, the deep neural network technology for speech recognition will promote the human-like auditory information processing technology as well. Subsequent development of speech recognition technology focuses on the design of the network＇s structure and the training algorithms of deep neural networks that make it better fit our auditory systems. Finally, we briefly analyze the feasibility of human-like anti-noise processing and human-like auditory attention mechanism implemented by deep neural networks.