
第 38 卷 第 6 期
2017 年 6 月
自 动 化 仪 表
PROCESS AUTOMATION INSTRUMENTATION
Vol. 38 No. 6
June. 2017
修改稿收到日期:2017 - 03 - 16
基金项目:特殊环境机器人技术四川省重点实验室基金资助项目(13ZXTK07)
作者简介:吴兴铨(1992—),男,在读硕士研究生,主要从事语音识别、软件开发等技术的研究。 E - mail:304094795@ qq. com。
周金治(通信作者),男,硕士,副教授,主要从事计算机网络与物联网、智能家居、语音识别等技术的研究。
E - mail:zhoujinzhi@ swust. edu. cn。
基于改进小波变换的语音基音周期检测
吴兴铨
1,2
,周金治
1,2
(1. 西南科技大学信息工程学院,四川 绵阳 621010;
2. 西南科技大学特殊环境机器人技术四川省重点实验室,四川 绵阳 621010)
摘 要: 基音在许多方面都有比较广泛的应用,比如语音编码、语音识别、语音转换、音乐检索以及发声系统疾病诊断等。 针对目前
很多小波变换方法在测量基音周期时存在的准确度低、复杂度高、鲁棒性差等缺点,以及在带噪语音环境下,特别是在非平稳噪声下
比较难判断语音基音周期的问题,提出了一种基于改进小波变换的语音基音检测方法。 首先将每帧带噪信号进行预处理,提取出有
话段的信息,消除直流分量;然后在加窗分帧后先进行端点检测,滤波后再分帧;接着再利用小波分解后取低频系数重构信号;最后结
合四阶累积法对重构信号进行基音检测。 试验结果表明,该方法在不同带噪语音环境下和低信噪比条件下,提高了带噪语音基音检
测的准确性。 与传统的小波变换法相比,该方法鲁棒性好且计算复杂度低,有利于语音基音周期检测。
关键词: 带噪语音; 基音检测; 小波变换; 重构信号; 三电平中心削波; 端点检测; 信噪比
中图分类号: TH - 3;TP391. 4 文献标志码: A DOI:10. 16086 / j. cnki. issn1000 - 0380. 201706016
Speech Pitch Period Detection Based on Improved Wavelet Transform
WU Xingquan
1,2
,ZHOU Jinzhi
1,2
(1. School of Information Engineering,Southwest University of Science and Technology,Mianyang 621010,China;
2. Robot Technology for Special Environment Key Laboratory of Sichuan Province,
Southwest University of Science and Technology,Mianyang 621010,China)
Abstract: Pitch has been widely used in many areas,such as speech encoding,speech recognition,speech conversion,music
retrieval and diagnosis of diseases of vocal system. Aiming at the disadvantages of pitch period measurement based on wavelet
transform,e. g. ,low accuracy,high complexity and poor robustness,and the problem of the difficulty of judging the speech pitch
period under noisy speech environment,especially under non - stationary noising,the speech pitch period detection method based
on the improved wavelet transform is proposed. Firstly,each frame of noisy signal is preprocessed to extract the information with
speech segments and eliminate the DC component. Then,endpoint detection is conducted after window - added framing,and re -
framing after filtering; and then the low frequency coefficient reconstructed signal is extracted after using wavelet decomposition.
Finally,the pitch detection for the reconstructed signal is accomplished combining with the fourth order cumulative method. The
experimental results show that under different noisy speech environment and low SNR condition,the accuracy of noisy speech pitch
detection is improved. Comparing with the method of traditional wavelet transform,this method features better robustness and low
computation complexity,these are helpful in speech pitch period detection.
Keywords: Noisy speech; Pitch detection; Wavelet transform; Restructured signal; Three - level center clipping; Endpoint
detection;SNR
0 引言
基音是语音信号的重要参数,语音信号中的音高
由基音频率
[1]
得到。 基音广泛应用于各个领域,如语
音编码、语音识别、语音转换、音乐检索以及发声系统
疾病诊断等
[2]
。 近年来,基音检测方法层出不穷,其中
具有代表性的方法有自相关函数法、平均幅度差函数
法、阴阳估计法、基于滤波器的算法、倒谱分析法、多分
辨率法和离散小波变换法
[3]
等。 因为语音信号是一个
非平稳、非准周期的信号,会受到声道共振峰的影响,
万方数据