We present a study on detecting repeated speech segments in consecutive utterances in a user session of a dialogue application. Such repetition patterns often carry important information of a user's expectation from or frustration about the dialogue system and therefore provide a source of clues for the system to adjust its dialogue strategies when such repetitions occur in order to improve system performance. We propose a recursive dynamictime warping based pattern comparison algorithm with no fixed end-points to find similar parts within the two utterances, called original and correction utterances. Candidate reference patterns are generated from the correction utterance by an unsupervised segmentation scheme. When no prior information about the position of the repeated parts is used, each reference pattern is compared with the original utterance from the beginning to the end. Experiments are conducted on 190 utterances of spontaneous speech from simulated dialogue sessions. The proposed algorithm achieves detection and rejection rates of 80.3% and 85.2% for the repeated and non-repeated parts, respectively. When segmentation information provided by the recognizer in the dialog system is incorporated into preparing the reference patterns, the rates are increased to 93.8% and 91.1%.
Bibliographic reference. Cevik, Mert / Weng, Fuliang / Lee, Chin-Hui (2008): "Detection of repetitions in spontaneous speech in dialogue sessions", In INTERSPEECH-2008, 471-474.