Application And Research Of Extreme Learning Machine In Speech Emotion Recognition

With the rapid development of computer technology, the intelligence andhumanity of a computer become a new hot research topic. While the voice actsas the easiest way for human to communicate with the computer, it becomes akey factor in the implementation of the intelligence and humanity of a computer.This research focuses on the speech emotion recognition, including thefollowing three studies:First, the paper reviewed several common network models in speechemotion recognition in terms of their principles, advantages and disadvantages.The algorithm of extreme learning machine was identified and applied to speechemotion recognition because of its fast learning speed and excellentgeneralization performance. A neural network model with a generalized singlehidden layer feed forward was developed based on extreme learning machine,including ELM (basic extreme learning machine) and KELM (extreme learningmachine with kernel recognition model). The developed neural network wascompared with the SVM model. Their recognition results of three types ofemotion (angry, happy, and neutral) from two speech emotion databases (TYUTand EMO-DB) were analyzed.Second, an Im-ABC (improved artificial bee colony) algorithm was developed to optimize KELM parameters, which have an important influenceon the performance of the network. The improved artificial bee colonyalgorithm overcame its shortcoming of reducing population diversity and slowconvergence. The recognition experiment with four types of emotion (happy,angry, sad, and neutral) from EMO-DB database confirmed that the modelKELM with Im-ABC algorithm to optimized parameters are better than themodel SVM with Im-ABC algorithm to optimized parameters on both time andgeneralization performance.Finally, the selective integration extreme learning machine model wasproposed to solve the problem of instability of the basic extreme learningmachine network. This involves two steps. Step1: establish a Bagging extremeleaning machine network and out-of-bag samples; Step2: treat the individual ofartificial bee colony as the weight vector of bagging extreme leaning machinenetworks，and the fitness function as the generalization error of out-of-bagsamples with different weight vector of bagging machine learning network; usethe artificial bee colony algorithm to get the optimal weight vector, weeding outthe ELM classifier with weight value less than the threshold, and integrating therest ELM classifiers. The recognition experiment with four types of emotion(happy, angry, sad, and neutral) from EMO-DB database compared theselective integration extreme learning machine with other three models: ELM(the basic extreme learning machine), majority vote V-ELM, and integratedBagging-ELM model. The results showed that selective integration extreme learning machine network was superior to other three models on both thestability and generalization performance.