Keyword spotting for dialectal speech and Introduction of wav2vec2.0

Northern towns of the northeastern Tohoku region of Japan have folktales unique to each region. They are told by storytellers in local dialects which make them difficult for speakers of standard Japanese to understand. We have proposed a system to detect keywords when they are spoken, and to display an explanation so that anyone can understand the tale contents. Keyword spotting (KWS) can be used with spoken keywords, but it requires the creation of dialectal spoken keyword. Therefore, we propose a method by which a keyword is given as monophone or as a syllable series of text. Then a frame-level subword series is constructed and matched with input dialectal speech. Additionally, we introduce wav2vec2.0 to obtain the posterior probability vectors of monophone and syllable to improve the detection accuracy. First, we compare the proposed method with the Query-by-Example Spoken Term Detection (QbE-STD) method using spoken queries. Next, we compare the proposed method with our conventional method for text queries, which reveals that the proposed method achieves a 4.37 percentage point (%pt) improvement in detection accuracy, and a further 2.64%pt improvement using frame reduction. Furthermore, the introduction of wav2vec2.0 improved the detection accuracy by 5.14%pt.