Can Automatic Speech Recognition Learn More from Human Speech Perception?

Although a great deal of progress has been made during the last two decades in automatic speech recognition (ASR), the performance of these ASR systems, as measured by word recognition and concept understanding error rates, is still much worse than that achieved by humans, even for carefully read and articulated speech in quiet conditions. This performance gap (between machines and humans) increases even more in noisy conditions and for conversational speech. Steadily increasing computational speed and computer memory tend to impose fewer and fewer constraints on the types and the amount of recognition processing that can be brought to bear on a particular recognition task. In spite of the increased computation and memory, the state-of-the-art technology in automatic speech recognition appears to have reached a plateau in the past few years. New techniques and principles need to be invented or applied in order to substantially reduce the current performance gap in speech recognition between humans and machines. This paper presents some ideas intended to stimulate further research on applying knowledge and principles derived from studies of human speech perception to automatic speech recognition. Although the mechanisms of human speech perception (HSP) are not fully understood, some findings from neuroscience, physiology, cognitive science and psychology could potentially lead to new understanding and thereby stimulate the development of new techniques and architectures for automatic speech recognition that, eventually, will bridge and reduce the performance gap between machines and humans.