Robustness of Automatic Speech Recognition
Most speech recognition research has shifted to conversational and natural speech in order to make effective and intuitive speech-enabled interfaces. Despite significant advances, many challenges remain to achieve the realization of efficient conversational systems. The ultimate goal consists of making ASR indistinguishable from the human understanding system. This chapter addresses the ASR robustness problem. It begins by giving the statistical formalism of speech recognition and then it describes the main robust features representing the hearing/perception knowledge that are used in the subsequent chapters. The major approaches used to achieve noise robustness are described. The relationship between dialog management systems and ASR is also investigated. Finally, a new paradigm giving soft computing techniques a new role in speech interactive systems is presented.