计算机工程 ›› 2018, Vol. 44 ›› Issue (5): 256-261.doi: 10.19678/j.issn.1000-3428.0046903

• 多媒体技术及应用 • 上一篇    下一篇

基于非负矩阵分解的情感语音基频转换研究

邓叶勋,赵晖   

  1. 新疆大学 信息科学与工程学院,乌鲁木齐 830046
  • 收稿日期:2017-04-20 出版日期:2018-05-15 发布日期:2018-05-15
  • 作者简介:邓叶勋(1991—),男,硕士研究生,主研方向为情感语音分析与转换、人工智能;赵晖(通信作者),教授、博士生导师。
  • 基金项目:
    国家自然科学基金(61561047)。

Research on F0 Conversion of Emotional Voice Based on Non-negative Matrix Factorization

DENG Yexun,ZHAO Hui   

  1. College of Information Science and Engineering,Xinjiang University,Urumqi 830046,China
  • Received:2017-04-20 Online:2018-05-15 Published:2018-05-15

摘要: 为解决情感语音基频转换过程中基频建模的间断性问题,提高生成语音的情感自然度,利用非负矩阵分解(NMF)技术,提出带有参数控制的情感语音基频转换方法。选择连续小波变换参数化基频并对语音韵律结构中的各层级进行独立建模,采用NMF将基频特征数据分解为基范例及其对应的权重,将目标基范例替换待转换语音基范例并重建目标语音基频。此外,引入激活度调整因子作为参数控制对现有模型进行优化。实验结果表明,在小数据库语料中,该方法在基频重建误差与情感力度方面都显示出优势,且能够有效地将中性语音转换为情感语音。

关键词: 情感语音转换, 连续小波变换, 非负矩阵分解, 基频转换, 韵律层级

Abstract: In order to solve the discontinuous problem of the F0 modeling in the process of the emotional voice conversion,and improve the emotional naturalness of the generated voice,a method for F0 conversion of emotional voice with parameter control based on Non-negative Matrix Factorization(NMF) is proposed.The Continuous Wavelet Transform (CWT) is used to parameterize F0 and model the different levels in the phonetic prosody structure.Then,the characteristic data of F0 is decomposed to the base exemplars and their weights by using the NMF,and by replacing the base exemplars of being converted voice with target,the F0 of the target voice is constructed.In addition,the activation factor,as control parameter,is introduced to optimize the existing model.Experimental results show that,this proposed method has a certain advantage in both the fundamental frequency reconstruction error and the emotional intensity,and can effectively convert the neutral voice to the emotional voice.

Key words: emotional voice conversion, Continuous Wavelet Transform(CWT), Non-negative Matrix Factorization(NMF), F0 conversion, prosody level

中图分类号: