基于GMM和Pitch-MFCC联合特征的语音转换  

Speech Conversion Based on GMM and Pitch-MFCC Joint Features

作　　者：任艺昊 陈汝真 张少华 余仁杰 邢沛然 REN Yihao;CHEN Ruzhen;ZHANG Shaohua;YU Renjie;XING Peiran(School of Information Science and Technology,Tibet University,Lhasa 850000,China)

机构地区：[1]西藏大学信息科学技术学院,西藏拉萨850000

出　　处：《系统仿真技术》2024年第4期329-333,共5页System Simulation Technology

摘　　要：在语音转换系统中,特征参数是语音转换的关键。大部分语音转换方法通常是提取源语音和目标语音的某一种特征参数,对其进行训练和合成。然而不同的特征参数具有不同的物理和声学意义,因此传统的语音转换方法可能会忽略不同特征之间存在的互补性,可通过对各特征参数进行联合构建有效解决此问题。为此,提出音高(Pitch)与梅尔频率倒谱系数(MFCC)相融合的Pitch-MFCC联合特征参数,并结合高斯混合模型(GMM)构建语音转换方法。通过仿真对所提出的方法进行验证,与基于GMM和MFCC的语音转换方法进行对比,结果表明,本研究提出的语音转换方法可提高语音转换质量。Feature parameters are the key to speech conversion in speech conversion system.For most speech conversion methods,only a specific feature parameter of source speech and target speech is usually extracted,trained and synthesized to generate the converted speech.However,different feature parameters have different physical and acoustic meanings,so the complementarity between different features may be ignored.This problem can be effectively solved by the method of joint construction of different feature parameters.On this basis,a speech conversion technology is proposed by combining the joint feature parameters of integrating Pitch and MFCC with Gaussian mixture model.The proposed method is verified by simulation,and compared with the methods of GMM and MFCC respectively based.Experimental results show that the proposed method can improve the quality of speech conversion.

关 键 词：语音转换 联合特征提取 GMM Pitch-MFCC 

分 类 号：TN912.3[电子电信—通信与信息系统]