计算机工程2025，Vol.51Issue(2)：35-53,19.DOI:10.19678/j.issn.1000-3428.0068739
低资源环境下的语音唤醒研究综述
Review of Research on Keyword Spotting in Low-Resource Environments
摘要
Abstract
Keyword Spotting(KWS)is a crucial technology for enabling human-computer interaction and has long been a focal point in speech technology research.As deep learning technology has advanced,research methodologies have transitioned from traditional Large-Vocabulary Continuous Speech Recognition(LVCSR)techniques to neural network-based approaches.However,challenges remain in achieving efficient KWS on small devices and training models with limited sample data,particularly in the design of low-resource KWS systems.This review begins by defining the concept of low-resource KWS,distinguishing it from general speech recognition and related terminology.It then introduces classic KWS models and their applicable scenarios while detailing the current global state of research on low-resource KWS.Next,mainstream technologies and optimization strategies for acoustic feature extraction and modeling are explained,with a focus on the structural components of KWS systems.An analysis of model lightweight methods is then conducted,where their advantages and disadvantages are compared.Common solutions for low-resource KWS,such as few-and zero-shot learning as well as transfer learning,are summarized,and common KWS datasets and evaluation metrics are introduced.Finally,future research directions for low-resource KWS technology are discussed.关键词
语音唤醒/低资源/模型量化/小样本学习/人机交互Key words
Keyword Spotting(KWS)/low-resource/model quantization/few-shot learning/human-computer interaction分类
计算机与自动化引用本文复制引用
王月昊,周若华..低资源环境下的语音唤醒研究综述[J].计算机工程,2025,51(2):35-53,19.基金项目
国家自然科学基金(11590774). （11590774）
