CAJ | 학술논문

关键词检测是从连续语音流中检测预先定义的给定词的技术，是语音识别领域的一个重要应用。目前的关键词检测研究中，主流的方法是基于连续语音识别器的先识别后检测的两阶段方法，语音识别器的准确率对关键词检测有很大影响。本文首先在识别阶段引入深度学习技术来改善关键词检测算法的性能。进而针对识别阶段和检测阶段缺乏紧密联系，耦合度不够的问题，研究了侧重关键词的深度神经网络声学建模技术，利用非均匀的最小分类错误准则来调整深度神经网络声学建模中的参数，并利用AdaBoost算法来动态调整声学建模中的关键词权重。结果表明，利用非均匀最小分类错误准则来调整深度神经网络参数进行优化的声学模型，可以提高关键词检测的性能。
관건사검측시종련속어음류중검측예선정의적급정사적기술，시어음식별영역적일개중요응용。목전적관건사검측연구중，주류적방법시기우련속어음식별기적선식별후검측적량계단방법，어음식별기적준학솔대관건사검측유흔대영향。본문수선재식별계단인입심도학습기술래개선관건사검측산법적성능。진이침대식별계단화검측계단결핍긴밀련계，우합도불구적문제，연구료측중관건사적심도신경망락성학건모기술，이용비균균적최소분류착오준칙래조정심도신경망락성학건모중적삼수，병이용AdaBoost산법래동태조정성학건모중적관건사권중。결과표명，이용비균균최소분류착오준칙래조정심도신경망락삼수진행우화적성학모형，가이제고관건사검측적성능。
Spoken term detection ( STD) is a task to automatically detect a set of keywords in continuous speech, which is an important field of speech recognition.Current study is based on two-stage approach i.e.recognition and detection.The accuracy of speech recognition has a significant impact on keyword detection.Firstly, this paper uses deep leaning tech-niques to improve performance during the first stage.As the two stages lack of close contact, the paper studies using non-uniform misclassification error ( MCE) criteria to adjust the parameters in deep neural network based acoustic modeling. Further the paper uses the adaptive boosting ( AdaBoost) strategy to adjust keywords’ weight dynamically.It shows that non-uniform MCE can improve the performance of STD.