摘要
儿向语音对早期儿童成长有较大影响,正确检测并充分利用儿向语音具有现实意义。为此,构建一种基于Adaboost算法的汉语儿向语音检测模型,以提高检测准确率。使用决策树作为弱分类器对提取的汉语儿向语音特征进行学习,并组成弱分类器元组,同时对该弱分类器组的分类结果进行加权,区分待测语音的类别。实验结果表明,汉语儿向语音的元音持续时长超过非儿向语音的元音持续时长;提升弱分类器的数量可提高汉语儿向语音检测正确率;分段语音时间越长,汉语儿向语音检测正确率越高;采用改进的Adaboost算法比采用v-SVM算法具有更高的准确率和精度,同时可增强系统的鲁棒性。
Child-directed Speech(CDS) has large influence on early child growth, so it is significant to recognize CDS from a speech and make full use of it. In order to improve the detection accurary, this paper constructs a Chinese CDS detection model based on Adaboost algorithm. It uses a decision tree as a weak classifier for extracting features of Chinese CDS to learn, and forms weak classifier tuple, while the classification results of this group of weak classifiers are weighted voting,to distinguish the voice category. Experimental results show that the vowel duration of CDS is longer than non- CDS ;increasing the number of weak classifiers will improve the accuracy of Chinese CDS ;the longer the length of test speech is,the higher the detection accuracy of Chinese CDS;compared with v-SVM algorithm, Adaboost algorithm has higher accuracy and precision in Chinese CDS detection,and it improves the robustness of the detection system.
出处
《计算机工程》
CAS
CSCD
北大核心
2017年第5期286-289,293,共5页
Computer Engineering
基金
国家自然科学基金面上项目(61173106)
关键词
儿向语音
语音检测
特征提取
ADABOOST算法
决策树
Child-directed Speech (CDS)
speech detection
feature extraction
Adaboost algorithm
decision tree