A new method to display the accuracy of results for each keyword of the KWS model

Nguyen Tuan Anh, Nguyen Thi Hang

Speech Keyword Spotting; KWS; Keyword Accuracy; Keyword Spotting Accuracy; KWA; Speech Recognition.

This study solves the accuracy problem of each keyword when training Keyword spotting (KWS) in non-aligned string results. This approach is called Keyword Detection Accuracy (KWA), which has been improved from the Levenshtein Distance algorithm, it is used to evaluate the accuracy of keywords in KWS by measuring the minimum distance between two strings. The main improvement algorithm is to display the status of each keyword in the training phase for predictive and true labels. In this study, the model used for training is LIS-Net, which is used in Speech Command Recognition. The results of the model are significantly improved compared to baseline models, and the results are displayed on graphs that can see the accuracy of each keyword.