Research And Implementation Of Laughter Detection In Continuous Speech

In recent years,speech recognition related technologies are becoming mature.However,our daily communication not only contains basic information expressed by the speaker,but also conveys the speaker’s emotional state,chatting atmosphere,etc..At present,more and more researchers focus on the study of these aspects,namely voice sub-information（paralinguistic）^[1]Research.In our daily natural speech,there are yawn,filling sound,applause,laughter and other non-text-based speech factors.To some extent,they reflect the atmosphere and background of the speech,the speaker’s emotional state,personality and other information.Therefore,researches on non-text speech are helpful to speech sub-information research.This thesis focuses on the laughter detection in continuous speech.Based on speech frame,ELM（Extreme Learning Machine）is used to detect laughter in continuous speech.Based on the continuous characteristics and duration characteristics in laughter,the test results in ELM algorithm are optimized.The main work is as follows:1.Detailed description and analysis are made in the structure and classification of laughter.Furthermore,laughter detection in continuous speech based on frame level is proposed.2.In order to address the level of detection frame based on laughter and the emergence of large-scale data,this paper is based on continuous speech ELM algorithm in laughter detection algorithm.With the expansion of the scale of the training data,the identify performance is gradually improved,and the operation rate of the system is maintained at second level,with respect to the previously more common SVM（Support Vector Machine）algorithms,the system’s performance has greatly improved.3.Based on the analysis of the continuity and duration of laughter event,the paper proposes an auxiliary judgment method based on the method of voting rules.Considering the continuity of laughter,to make a frame classification,the article takes the frame to the center,from left to right,each side has 20 frames,a total of 41 voting results will be the final recognition result,which makes the value F in the system increased by 8.49%and reached F=80.27%.