Alexander Rogath Kivaisi;Qingjie Zhao;Yuanbing Zou
{"title":"Multi-Task ConvMixer Networks with Triplet Attention for Low-Resource Keyword Spotting","authors":"Alexander Rogath Kivaisi;Qingjie Zhao;Yuanbing Zou","doi":"10.26599/TST.2024.9010088","DOIUrl":null,"url":null,"abstract":"Customized keyword spotting needs to adapt quickly to small user samples. Current methods primarily solve the problem under moderate noise conditions. Recent work increases the level of difficulty in detecting keywords by introducing keyword interference. However, the current solution has been explored on large models with many parameters, making it unsuitable for deployment on small devices. When applying the current solution to lightweight models with minimal training data, the performance degrades compared to the baseline model. Therefore, we propose a light-weight multi-task architecture (<9.0×10>4</sup>\nparameters) created from integrating the triplet attention module in the ConvMixer networks and a new auxiliary mixed labeling encoding to address the challenge. The results of our experiment show that the proposed model outperforms similar light-weight models for keyword spotting, with accuracy gains ranging from 0.73% to 2.95% for a clean set and from 2.01% to 3.37% for a mixed set under different scales of training set. Furthermore, our model shows its robustness in different low-resource language datasets while converging faster.","PeriodicalId":48690,"journal":{"name":"Tsinghua Science and Technology","volume":"30 2","pages":"875-893"},"PeriodicalIF":6.6000,"publicationDate":"2024-09-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=10691379","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Tsinghua Science and Technology","FirstCategoryId":"94","ListUrlMain":"https://ieeexplore.ieee.org/document/10691379/","RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"Multidisciplinary","Score":null,"Total":0}
引用次数: 0
Abstract
Customized keyword spotting needs to adapt quickly to small user samples. Current methods primarily solve the problem under moderate noise conditions. Recent work increases the level of difficulty in detecting keywords by introducing keyword interference. However, the current solution has been explored on large models with many parameters, making it unsuitable for deployment on small devices. When applying the current solution to lightweight models with minimal training data, the performance degrades compared to the baseline model. Therefore, we propose a light-weight multi-task architecture (<9.0×10>4
parameters) created from integrating the triplet attention module in the ConvMixer networks and a new auxiliary mixed labeling encoding to address the challenge. The results of our experiment show that the proposed model outperforms similar light-weight models for keyword spotting, with accuracy gains ranging from 0.73% to 2.95% for a clean set and from 2.01% to 3.37% for a mixed set under different scales of training set. Furthermore, our model shows its robustness in different low-resource language datasets while converging faster.
期刊介绍:
Tsinghua Science and Technology (Tsinghua Sci Technol) started publication in 1996. It is an international academic journal sponsored by Tsinghua University and is published bimonthly. This journal aims at presenting the up-to-date scientific achievements in computer science, electronic engineering, and other IT fields. Contributions all over the world are welcome.