In order to effectively utilize non-lexical audio events to analyze the semantics of conversational speech,the characteristic differences among the audio events frequently occurring in spontaneous vocalization are analyzed,and a two-stage decision-based method to detect non-lexical audio events in spontaneous vocalization is proposed.In this method,the characteristics of audio events are used to construct signal segments of audio events： the thre-shold decision is used to detect longer applause（the first-stage decision）,and statistical models are employed to detect other audio events（the second-stage decision）.Experimental results show that the average precision,the recall rate and the F1-measure of the proposed method for three non-lexical audio events（i.e.filled pause,laughter and applause） are respectively 87.3%,93.8% and 90.4%;and that,as compared with the existing method,the proposed method is of an average F1-measure increase by 7.5% and it helps to determine the boundaries of non-lexical audio events with higher accuracy.