Telephone Channel Natural Voice Keywords Detection

As one special field in speech recognition research, keyword spotting is to determine occurrences of one or more keywords embedded in unconstrained extraneous speech and/or noise. It has bright future in many application areas. In this paper, we give a brief history of keyword spotting research and provide a discussion of its fundamental principle in which three most important problems in this field are pointed out, that is, how to extract and choose feature and how to characterize keywords and garbage; and how to detect keywords from continuous speech. This paper describes the basic theory of HMM and presents simple and practical methods for building HMM and time aligning patterns with models are provided, that is, Segmental k-means training algorithm and Frame-synchronous Viterbi algorithm. And we build a recognition-verification system which can detect keywords from continuous speech without grammar restriction. Furthermore, we do some works in feature transformation, speaker clustering, speaker adaption and Gaussian selection for improving system robustness and efficiency. Feature transformation is achieved by FastICA algorithm. Speaker clustering is implemented by Gaussian Mixture Model which can make system applied to a wider group of people and speaker adaption is achieved by Maximum Likelihood Linear Regression algorithm. And we use Gaussian selection method to reduce calculation. In utterance verification phase, some new confidence measures based on recognition results' information are used to reduce the false alarm rate. Finally, the paper shows the further research direction in this field.