Automatic Turn Segmentation in Spoken Conversations

In this paper we have studied the problem of detecting the spoken turn boundaries in human-human spoken conversations. The automation of this task is essential to enable the analysis, recognition and understanding of the speech transcriptions and dialog structures (e.g. turn taking, dialog act segmentation etc.). The problem formulation is different from previous work on metadata extraction in that we work on the time domain for the detection of boundaries. This approach has the advantage of giving fine grain measures of speech events and does not rely on the automatic speech transcriptions. We have explored applicability of different algorithms for this task and have found that a hidden Markov model combining results of the modulation spectrum analysis and Kullback-Leibler divergence of adjacent signal portions produces the best results. The performance of the algorithms has been evaluated on the Switchboard conversational speech corpus.