Building the next generation of computational psycholinguistic models of speech perception

点击翻译按钮获取中文摘要

How do human beings perceive and understand speech so effortlessly? Our brains give us the illusion that the process is simple, but the often strange errors made by even the most advanced artificial intelligence systems, and the baffling difficulty we have even taking stock of what we hear when we hear an unfamiliar language are clear clues that the process is in fact not easy at all. Incredibly, infants' perception becomes specialized in the language(s) they hear at home well before they can speak: as early as six months, a time when experiments also demonstrate that they recognize and understand the meaning of dozens or hundreds of common words. The cognitive science of speech perception has greatly advanced our understanding of how human speech perception works, and, to a lesser extent, of how the ability develops in infants. Nevertheless, our understanding is still very far from being advanced enough to build "human-like" computer systems that learn and perceive speech, and our current speech technology, tuned on implausibly large quantities of data, behave in many ways very differently from human beings. We seek to take advantage of recent advances in machine learning and speech technology to advance our understanding of (1) learning: what kind of systems can do the work of the infant brain and autonomously learn to decode the speech signals they hear into individual consonant and vowel sounds (currently we know of none)? do these systems end up making the same kinds of misperception errors as human listeners? (2) the early stages of auditory processing: many new speech processing systems appear, at first glance, to behave much more like the human auditory system than previous generations of speech technology, but further experiments with human listeners are needed to assess this, and to understand the implications for our understanding of human auditory processing if it is true; and, (3), how speech sounds are encoded by the brain in our memory for words. The answers to these questions have consequences for our understanding of how humans decode speech, how we learn to do this at an early age, and how we can build artificial intelligence systems that are less fragile, and that are capable of operating in far more of the world's languages.