Research On Utterance Verification Of Keyword Spotting Technology

Keyword spotting is an important research field in speech recognition. It not only has a better flexibility but also a high application value in the scope of the speech recognition. In many automatic speech recognition applications, it often needs to provide a confirmed identification mechanism to verify the correctness of some assumptions. The utterance verification is the process which verifies the reliability from the result of the speech recognizer. In this paper, we research on several key words detection technology and utterance verification technologies, which have a wide application.We focus on the detection strategies and recognition method of the keyword recognition system. Under the research on the detection technology of the garbage model, we proposal a method used to define to competition set in continuous speech which backtrack the effective relative path in constant time, then compare with the real voice clips to determine whether each particular segment should belong to a competitive set. In viterbi search, we use one-pass strategy which saves the store space, improves the recognized accuracy and can also use the utterance verification to correct some recognition error comparing to the two-pass strategy. A novel approach that is the weighted likelihood ratio fusion based on the revised rival-model is proposed. The algorithm distinguishes between different competitive set and effectively improve the recognition effect. Finally, we combine the VQ and SVM technology into utterance verification which not only fully use the advantages of classification in SVM and relatively fast computation speed in VQ but also avoids the huge computing volume when classify the SVM template of the speech. Through a series of experiments, we demonstrate that the proposed method is superior to other methods to some extent.