{"title":"An End-to-End Model Based on TDNN-BiGRU for Keyword Spotting","authors":"Shuzhou Chai, Zhenye Yang, Changsheng Lv, Weiqiang Zhang","doi":"10.1109/IALP48816.2019.9037714","DOIUrl":null,"url":null,"abstract":"In this paper, we proposed a neural network architecture based on Time-Delay Neural Network (TDNN)Bidirectional Gated Recurrent Unit (BiGRU) for small-footprint keyWord spotting. Our model consists of three parts: TDNN, BiGRU and Attention Mechanism. TDNN models the time information and BiGRU extracts the hidden layer features of the audio. The attention mechanism generates a vector of fixed length with hidden layer features. The system generates the final score through vector linear transformation and softmax function. We explored the step size and unit size of TDNN and two attention mechanisms. Our model has achieved a true positive rate of 99.63% at a 5% false positive rate.","PeriodicalId":208066,"journal":{"name":"2019 International Conference on Asian Language Processing (IALP)","volume":"521 1","pages":"0"},"PeriodicalIF":0.0000,"publicationDate":"2019-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"2","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"2019 International Conference on Asian Language Processing (IALP)","FirstCategoryId":"1085","ListUrlMain":"https://doi.org/10.1109/IALP48816.2019.9037714","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"","JCRName":"","Score":null,"Total":0}
引用次数: 2
Abstract
In this paper, we proposed a neural network architecture based on Time-Delay Neural Network (TDNN)Bidirectional Gated Recurrent Unit (BiGRU) for small-footprint keyWord spotting. Our model consists of three parts: TDNN, BiGRU and Attention Mechanism. TDNN models the time information and BiGRU extracts the hidden layer features of the audio. The attention mechanism generates a vector of fixed length with hidden layer features. The system generates the final score through vector linear transformation and softmax function. We explored the step size and unit size of TDNN and two attention mechanisms. Our model has achieved a true positive rate of 99.63% at a 5% false positive rate.