After many years' development, Automatic Speech Recognition (ASR) technology has been mature; and many ASR systems have been developed. However, there are still many challenges confronting researchers, such as noise reducing, environment adaptation and Out-Of-Vocabulary (OOV) words detection and rejection etc. It is important to solve these problems to enhance the robustness and reliability of ASR systems. In this paper we briefly look back on stochastic speech recognition fundamentals and the framework of ASR systems. Then we focus on the rejection problem in ASR. Our discussion of rejection problem ranges from its definition, essence and evaluation to the popular solutions nowadays. We note that researchers Usually solve it in two ways, namely explicit modeling and implicit modeling (mainly by confidence measure). Due to the constraints of specific ASR tasks and specific recognizer to rejection, we propose several platform-dependent rejection approaches in two ASR platforms which models whole word and mandarin tri-phone respectively. In addition, we try some platform-independent rejection algorithms based on Syllable Segmental Models (SSM). The application of ASR technology in embedded systems is an important direction for its popularization. The development of Embedded Speech Recognition Systems (ESRS) involves in many factors such as hardware and software, resources and performances etc. In this paper, we analyze and expound the main problems of ESRS in details and give several examples we made during postgraduate stage. Finally we discuss the promising prospect of ASR SOC.
修改评论