Discriminatively Derived Hmm-Based Announcement Modeling Approach for Noise Control Avoiding the Problem of False Alarms

摘要

Earlier we proposed modeling echo residuals by using multiple echo models built from a set of specific amouncement. Experienced callers may interrupt the prompt by speaking the keywords over the prompt. This leads to incomplete prompt echoes that was not properly modeled by multiple echo models. In this study, we investigate further improvements by building an echo model of each word in the entire announcement, then linking each model in sequence to track the exact echo that precedes valid speech (movie title). The experimetnal results show that by modeling exactly, one can get better recognition accuracy and less false triggering, with a possible increase in computational complexity.