Two-level approach for detecting non-lexical audio events in spontaneous speech

Based on analyses of characteristic differences between various audio events, a two-level approach is proposed for detecting three non-lexical audio events (filled pause, laugh, and applause) in spontaneous odel-based decision. The experiments give average precision of 87.3%, recall of 93.77%, and F-measure of 90.42%. Compared with the sliding window based approach, average F-measure is improved by 7.52%. Moreover, it can more accurately determine the boundaries of non-lexical audio events in spontaneous speech.