Research And Realization Of Voice Activity Detection Based On Multiple Observation Likelihood Ratio System

Voice Activity Detection is the technology which can extract the startingand ending points from speech signal. Since voice activity detection plays animportant role in many technologies such as speech analysis, speech synthesis,speech encoding and speech recognition, this paper researches on the voiceactivity detection technology, and realizes a voice activity detection system.Because there have been many voice activity detection methods, andmultiple observation likelihood ratio based voice activity detection is simple andcan achieve good performance, this paper carrys out voice activity detectionutilizinge multiple observation likelihood ratio feature. First we divide timeframes from the speech signal, then we estimate the noise from the signal, andcompute the likelihood ratio of each frame under the result of noise estimation.We constitute the features with the likelihood ratios of this frame and certainnumbers of previous frames.After feature extraction, we train the decision rule through the trainingsamples. This paper utilizes linear classifier as the decision rule model. S o theproblem is how to train the proper weights of linear classifier to make thedetection as accurate as possible. This paper presents Minimum ClassificationError model, Maximum Area Under the ROC Curve model and Support VectorMachine model, and proposes an extended MaxAUC model. The experimentalresults indicate that the extended MaxAUC model and SVM model have betterstability and detection performance.In order to understand the real requirement and module function, this paperanalyzes the requirement and designs this system. In requirement analysis, thispaper lists the demands of this system, and decompose this system as signalinput, feature extraction, training and testing. So as to present the relationshipsbetween those modules, this paper uses data flow diagram to portray thetransmition and processing of data in this system. In system designation, thispaper presents the working process of this system. For the important modules inthis system, this paper describes their functions, workflows, parameters and return values. The requirement analysis and system designation lay thefoundation for the subsequent system realization and system testing.Finally, in order to verify the performance of this system, this paper teststhe system. This paper respectively tests the noise estimation and each decisionrule models. The results show that this system can achieve good performance.