Keyword spotting result post-processing to reduce false alarms
Šmídl Luboš, Jan Trmal · 2009
This paper proposes a new technique to reduce false alarms in a keyword spotting system. HMM based acoustic keyword spotting system can be easily set to get the desired detection rate and the corresponding false alarm rate by thresholding value based on ROC (receive operating curve) characteristic. Unfortunately, as practical experiences show, inexperienced users have problem to set the threshold correctly. The presented techniques describe reduction of false alarms in one order of magnitude without any decrease of recognition accuracy.