A Persian Wake Word Detection System Based on the Fine Tuning of A Universal Phone Decoder and Levenshtein Distance

This paper deals with the preparation of a universal phone decoder and using it to build a speech-based Persian wake word detection system. First, two sets of spoken data in Persian are used to adjust the symbols and fine-tune the parameters of the phone decoder, which is named Allosaurus, until a wake word detection system with high accuracy is obtained. During this process, a slightly modified version of the Levenshtein Distance algorithm is used to calculate a confidence score for the system output decision. After the initial wake word detector is ready, the values used for calculating the Levenshtein Distance and their weights are optimized in order to achieve the highest possible accuracy. In the end, this work focuses on also maximizing the accuracy of noisy speech signal inputs, which is something that hasn’t been done in previous works.