Automatic speech recognition for mobile and personal applications

点击翻译按钮获取中文摘要

In this proposal, we examine ways to improve the performance of automatic speech recognition (ASR), whichconverts ordinary speech into text. This allows unsupervised programs to interact with people via voice, thusgiving people access to a broad range of data services, over telephones or PDAs (personal digital assistants).ASR exists now only for limited services, as its accuracy is much lower than that of human listeners. For ASRto gain widespread use, we need to make it much more efficient. This proposal addresses several areas oftargeted interest, e.g., accessing databases over wireless channels. The research deals with intelligent signalprocessing, as we need to develop algorithms to interpret voice signals in terms of their intent.The simplicity of models used in current ASR, when compared to the complex, nonlinear processing done byhumans, suggests that there remain many ways to improve ASR. At each level (i.e., feature estimation,temporal and acoustic modeling, language modeling, decision making), compromises have been made in ASRto have simple and fast processing, at the expense of lower recognition accuracy. We intend to examineseveral of these compromises and alternative ways to approach the performance of human speech recognition.A major application of our work is to render search and access to information via voice, e.g., large databasesas found on the World Wide Web. Typing on small portable devices is awkward; thus access via voice ishighly desired.Human-computer interactions often use a mouse+keyboard as machine input, and a computer screen orprinter as output. Speech, however, has always had a high priority in human communication, developed longbefore writing. While less economical in communication bandwidth than text, speech is more universal andeasy to use than text processing, e.g., people speak more quickly and easily than typing. The proposed researchhere deals with improving ASR to render it more practical and hence more useful.On recherche nouvelles méthodes afin d'améliorer la performance de systèmes de reconnaissance automatiquede parole (RAP), ce qui convertit la parole ordinaire en texte. Ceci permet aux algorithmes non dirigés àinteragir avec les gens par la parole, ce qui leur donne accès à une grande gamme de services de données, partéléphone ou par assistants numériques.