Prediction of keyword spotting accuracy based on simulation

This paper proposes a method of predicting accuracy of keyword spotting in terms of FA count and spotting score of correct detections. A new measure F for predicting the FA count is calculated by simulation of the keyword spotting for phoneme sequences that phoneme-based language model generates. Another measure C for predicting the spotting score of correct detections is obtained from a product of correct recognition probabilities of phonemes. Both correlation coe cients and prediction errors are used to evaluate these measures in comparison with a simple measure of the keyword phoneme length, L. The prediction errors of FA count based on L was 7.71. The measure F reduced the prediction errors by 16%, and it had stronger correlation with the FA count. Furthermore a combined measure of F and L reduced the errors by 23%. On the other hand, L was more e ective to predict the spotting score of correct detections than the measure C.