KWA: A New Method of Calculation and Representation Accuracy for Speech Keyword Spotting in String Results

Abstract

This study proposes a new method to measure and represent accuracy for Keyword Spotting (KWS) problem in nonaligned string results. Our approach, called Keyword Spotting Accuracy (KWA), was improved from the Levenshtein Distance algorithm, that used to evaluate the accuracy of the keywords in KWS by measuring the minimum distance between two strings. The main improved algorithm is to show the status of each keyword in training phase for predicted and true labels. In which, representing which words are correct, which ones need to be inserted, substituted or deleted when comparing the prediction labels with true ones during the training phase. In addition, a new method of presenting the multiple keywords in results was proposed to indicate the accuracy of each keyword. This method can display detailed results by keywords, from which, we can obtain the accuracy, distribution, and balance of the keywords in the training dataset by actual speech variance, not by counting keywords in true labels as usual.