False alarm reduction by improved filler model and post-processing in speech keyword spotting

This paper proposes four methods for improving the performance of keyword spotting (KWS) systems. Keyword models are usually created by concatenating the phoneme HMMs and garbage models consist of all phonemes HMMs. We present the results of investigations involving the use of skips in states of keyword HMMs and we focus on improving the hit ratio; then for false alarm reduction in KWS we model the words that are similar to keywords and we create HMMs for highly frequent words. These models help to improve the performance of the filler model. Two post-processing steps based on phoneme and word probabilities are used on the results of KWS to reduce the false alarms. We evaluate the performance of the improved keyword spotting in FarsDat corpus and compare the approaches. The presented techniques depict better performances than the popular KWS systems.