计算机与数字工程2025，Vol.53Issue(3)：766-769,794,5.DOI:10.3969/j.issn.1672-9722.2025.03.026
轻量级实时语音唤醒词引擎研究
Research on Lightweight Real-Time Voice Wake-Up Word Engine
摘要
Abstract
The design and implementation of an accurately recognized wake lexicon is the basis of speech assistant implemen-tation,and the construction of the wake lexicon is determined by an efficient and reliable search engine model.This paper firstly es-tablishes the initial wake lexicon and the candidate wake lexicon,and represents the above two audio samples by logmel spectro-grams,designs a search engine consisting of the first four modules of EfficientNetb0 system,calculates the Euclidean distance be-tween the wake words in the candidate lexicon and the spectrograms in the initial wake lexicon,and transforms the similarity be-tween the wake words,and converts the words less than the specified threshold.The candidate words that are smaller than the speci-fied threshold are judged as new wakeup words and extended to the wakeup lexicon.The wake-up lexicon constructed by this engine can effectively improve the computational efficiency of the wake-up word engine with the guaranteed accuracy of 95.40%.关键词
唤醒词/EfficientNetb0体系/小样本学习/孪生神经网络/欧几里得距离Key words
wakeup words/EfficientNetb0 system/one-shot learning/euclidean distance/Euclidean distance分类
信息技术与安全科学引用本文复制引用
燕佳伟,张俊,年梅..轻量级实时语音唤醒词引擎研究[J].计算机与数字工程,2025,53(3):766-769,794,5.基金项目
国家重点研发计划子课题"多模态信息高效处理技术"(编号:E1182101)资助. （编号:E1182101）
