Research On Confidence Measure In Speech Keyword Recognition

Speech keyword recognition is to search given keywords in continuous and unlimitedspeech data. It has broad application prospects. Keyword recognition system consists of keyworddetection and confidence measure. In the phrase of keyword detection, the system generatesmaximum keyword hypotheses to improve the detection rate. Then, we should assess theconfidence of the candidate keywords to reduce the false alarm rate of the system. Therefore, theconfidence estimation of the candidate keywords is of crucial importance. As the false alarm rateof the keyword spotting system is high, effective confidence measures are investigated toenhance the discrimination of the confidence and improve the performance of the system.Research contributions of the thesis are listed as follows:A confidence measure based on frame-level sub-word posterior probability of Multi-layerPerception (MLP) is presented. Conventionally, the confidence is calculated from the acousticand language model scores computed by the recognizer of HMM model, which makes someincorrect assumptions, such as the frame-wise, possibly component-wise independence ofacoustic features and a finite number of Gaussian mixtures. The proposed confidence measure isdirectly calculated from the frame-level sub-word posterior probabilities produced by a MLPnetwork. The confidence estimation is completely separated from the keyword detection andthey use two different models. With this separation, decision making can be addressed with morereliable confidence and multiple confidence features can be integrated to improve the decisionquality. The experimental results show that the proposed approach is better than the mainstreamconfidence measures in the framework of HMM model and they have good complement, whencombining with the mainstream confidence measures in the scheme of HMM model, theperformance achieves further improvement.An improved confidence measure based on time and boundary feature is proposed. Themainstream posterior probability confidence measure underutilizes the phonetic pronunciationchanges and can not fully take into account the time and boundary feature of the candidate arcswhen we integrate the arcs with the same words which are overlapped in time. Themisrecognition results in the phrase of keyword detection can be easily introduced to theconfidence calculation and may miss some useful confidence information which can affect theeffect of confidence measure. To solve these problems, a relaxation rate is introduced to have aflexible selection of the segmental arcs which have the same words and the duration togetherwith the start and end boundaries satisfy the condition for the calculation of confidence. Theexperimental results show that the proposed approach is better than the mainstream lattice-basedposterior probability confidence measures and has good confidence performance.A discriminative confidence measure based on score remedy strategy of Support VectorMachine (SVM) is put forward. Whether the candidate keyword is correct or not can be treatedas a two-class classification problem and SVM is a good discriminative model with highclassification accuracy. The experimental results show that when the confidence score of thecandidate keyword is mapped to discriminative confidence, the discrimination of the confidence can be further enhanced. To train a SVM model, a posterior remedy approach is introduced tohandle the problem of data imbalance. Firstly, we have to estimate the class prior probabilityfrom sample data. Afterwards, the proposed approach is used to amend the class posteriorprobabilities from the SVM. Finally, the revised results are used as the confidence of thecandidate keywords. The experimental results demonstrate that the proposed posterior remedyapproach is effective.