Research And Implement Of Intelligent Approach For Digital Music Speech Evaluation

With the popularization and development of musical entertainment, the research of intelligent processing about digital music has attracted widespread attention. Scoring the singing by intonation and pitch has been applied in many musical entertainment software products. But the humming recognition and emotion recognition about music speech are still in infancy,we need to further expand the computer intelligent evaluation methods to solve these problems.Compared to normal speaking speech, the speech of digital music changes with the wide range of pronouncing,having great impact by songs’ rhythms and high requirement of real-time processing. The effect of processing digital music by traditional ways is not ideal,such as single word can not be cut apart accurately,the traditional ways to select features can not be effectively applied to music speech.Considering music speech’s feature of wide range of pronouncing,this paper proposes humming recognition algorithm based on information entropy and speech segment algorithm based on particle swarm optimization.Information entropy is very sensitive to humming’ non-stop feature when pronouncing,so that it can be used to detective the humming accurately when people cheat to get higher grade.Traditional endpoint detection techniques are difficult to adapt to digital music speech’s feature of fast-changing rhythm, irregular pronounce intervals and characteristics of no pause or no obvious pause between words.So this paper proposes a method that can subdivide the multiple words which are not be segmented at first. This method can intelligently determine the number of word contained by a long speech segment and find exactly each word’s starting&end positions.It can be applied on the condition with noise.Experiment results show that these two algorithms can handle the digital music voice effectively and accurately and improve the accuracy of music speech segmentation.There are not any certain characteristics that can be very typical, individual representation and distinguish between different emotions at present.So we need to select the appropriate features from the existing known features for the emotion recognition.This paper proposes a emotion recognition algorithm based on DTW coefficient and SVM voting mechanism. Considering the prosodic features of music,this method extracts the characteristics which have little similarity between classes and big similarity in the same class by DTW coefficient.Then we build some different dichotomous SVM classifiers to classify each sample.According to the classifying result,we can get the emotion recognition result.Through the compare with algorithm used Fisher criterion and hierarchical SVM,we can know clearly that the algorithm proposed by this paper has improved the recognition accuracy effectively because of eliminating the effect to projection area on the Fisher caused by different numbers of samples and the impact of cumulative errors caused by hierarchical algorithm.Finally,we package the above algorithm into modules to put into KTV evaluating system.This paper considers the different between digital music speech and normal speech, proposing humming recognition algorithm, music segmentation recognition algorithm and emotion recognition algorithm which have high research value and broad development prospects to make up the lack of traditional algorithms.