基于多分类支持向量机和主体延伸法的基音检测算法  

Pitch detection algorithm based on multi-class support vector machine and main body extension method

作　　者：冯起斌 李鸿燕 FENG Qibin;LI Hongyan(College of Information and Computer,Taiyuan University of Technology,Jinzhong 030600,China)

机构地区：[1]太原理工大学信息与计算机学院

出　　处：《现代电子技术》2019年第22期150-154,158,共6页Modern Electronics Technique

基　　金：山西省自然科学基金(201701D121058)~~

摘　　要：低信噪比环境下的基音检测颇具难度却极有现实意义,传统基音检测在此背景下效果不佳。因此,提出一种基于多分类支持向量机的基音检测算法。该算法使用语音信号的静态帧级特征对多分类支持向量机进行监督训练,计算出各帧语音可能的几个基音大小,作为对应的基音候选值,并使用主体延伸法对得到的候选基音状态进行处理,结合帧与帧之间的时序信息,在候选基音中选取合适值连接起来得到被测语音的基音状态估计曲线。将该算法与相关方法进行比较,实验结果表明,该方法有效提升了低信噪比环境下的基音检测率,在不同强度的噪声干扰下仍能保持良好的鲁棒性。Pitch detection in low signal to noise ratio(SNR)environment is very difficult but very meaningful.However,traditional pitch detection is not effective in this context.Therefore,a pitch detection algorithm based on multi class support vec tor machine is proposed,in which the static frame level features of speech signals are used for supervised training of multi classi fication support vector machines.The possible pitch values of speech in each frame are calculated as the corresponding pitch candidate values.The obtained candidate pitch state is processed by the main body extension method.In combination with the time sequence information between frames,the appropriate values in the candidate pitches are selected and connected to obtain the pitch state estimation curve of the tested speech.Experimental results show that,in comparison with the other related meth ods,the proposed method can effectively improve the pitch detection rate in the environment of low SNR and can still maintain good robustness in noise interference with different intensity.

关 键 词：基音检测 监督学习 基音候选值计算 多分类支持向量机 主体延伸法 检测验证 

分 类 号：TN912[电子电信—通信与信息系统] 34[电子电信—信息与通信工程]