Analysis of disfluent repetitions in spontaneous speech recognition

In this paper, we investigate the effect of disfluent repetitions in spontaneous speech recognition. We characterize the repetition errors in an automatic speech recognition framework using repetition word error rate (RWER). The problem is addressed by both building classifiers based on acoustic-prosodic features and a multiword model for modeling repetitions. We also analyze the repetition word error rate for different acoustic and language models in the Fisher conversational speech corpus. The classifier approach is not promising on recognizer output and generates a high degree of false alarms. The multiword approach to modeling the most frequent function word repetitions results in an absolute RWER reduction of 1.26% and a significant absolute WER reduction of 2.0% on already well trained acoustic and language models. This corresponds to a relative RWER improvement of 75.9%.