基于线性预测能量谱系数的自然语音和耳语音的分类研究  

Research on LPESC-based classification of natural speech and whispered speech

作　　者：姚真真 胡金瑶 艾斯卡尔·艾木都拉[1] YAO Zhenzhen;HU Jinyao;ASKAR Hamdulla(School of Information Science and Engineering,Xinjiang University,Urumqi 830046,China)

机构地区：[1]新疆大学信息科学与工程学院,新疆乌鲁木齐830046

出　　处：《现代电子技术》2023年第2期85-90,共6页Modern Electronics Technique

基　　金：国防科技基础加强计划(2021-JCJQ-JJ-0059);国家自然科学基金项目(U2003207)。

摘　　要：在进行语音识别之前对自然语音和耳语音进行预分类,再分别放入各自的识别系统,可以提高耳语语音识别系统的识别性能。基于此,文中提出一个新的特征线性预测能量谱系数(LPESC),在该特征的提取过程中,对提取到的频谱图进行切分,以便获取到更多的语音信息,并将其用于耳语音分类。此外,还特别设计4种不同的滤波器组,并将提取到的特征应用于7个分类器上。实验结果表明,密集的均匀三角滤波器组更加适合提取该特征,在7种传统分类器上均有较好的分类效果,其中SVM分类效果最好。最后,对比LPESC与传统特征(39维的LFCC和MFCC)在7种分类器上的分类效果,验证新特征的有效性。实验还发现,女生的耳语音有更好的分类效果。The recognition performance of the whispered speech recognition system can be improved by pre classifying natural speech and whispered speech before speech recognition,and then putting them into their respective recognition systems.A new feature linear prediction energy spectrum coefficient(LPESC) is proposed. In the process of feature extraction,the extracted spectrogram is segmented to obtain more speech information and apply it to whispered speech classification. In addition,4 different filter banks are designed specially,and the extracted features are applied to 7 classifiers. The experimental results show that the dense uniform triangular filter banks are more suitable for extraction of these features,and have good classification effects on 7 traditional classifiers,in which SVM is the best. The classification effects of LPESC and traditional features(39-dimensional LFCC and MFCC)on 7 classifiers are compared to verify the effectiveness of the new features. It is also found in the experiments that it has a better classification effect for girls ’ whispered speech.

关 键 词：语音分类 语音识别 耳语音 线性预测能量谱系数 特征提取 频谱图切分 结果分析 效果验证 

分 类 号：TN911.23-34[电子电信—通信与信息系统]