电讯技术
第
53
卷第
10
期
2013
年
10
月
Telecommunication Engineering
doi:
10.
3969/j.
issn.
1001-893x.
2013.
10.
014
Vol.
53
No.
10
Oct.
2013
基于压缩感知观测序列的语音信号基音周期提取＊
（南京邮电大学通信与信息工程学院，南京
210003)
摘
要：语音信号的清浊音辨别及基音周期的提耳又是语音信号处理的重要组成部分。提出了一种基
于压缩感知（
cs
）观测序列的语音基音周期才是取方法，大大减少了基音周期才是取的计算量。首先分
析不同观测矩阵下观测
l
序列的特征，选择了行阶梯矩阵作为所提方法的观测矩阵；然后对此观测矩
阵下的观测序列做小波分解，对小波分解的低频系数采用自相关方法提取基音周期。仿真与分析表
明，年口传统方法比较，所提方法与传统方法基音检测准确度相当，且在语音含有高斯白噪声的条件下
运行良好，但计算量仅约为传统方法的
1/160
关键词：语音信号处理；基音周期；压缩感知；非重构；自相关函数；小波变换
中图分类号：
TN912.
3
文献标志码：
A
文章编号：
1001-893X
(
2013)
10-1327-04
Pitch Detection Based on Compressed Sensing Sequence
CAO
Yang
(College
of
Telecommunication & lnfonnation Engineering,
Nanjing
University
of
Posts
and
Telecomrnunicalions,
Nanjing
2 l 0003, China)
Abstract:
Voiced/
unvoiced
decision
and
pitch
detection
technique
is
an
important
part
of
speech
signal
processing.
Based
on
the
observation
sequence
。
f
compressed
sensi
吨（
CS)
, a
method
to
detect
the
pitch
is
proposed
in
this
paper,
and
the
caculation
amount
is greatly
reduced
.
Firstly,
the
characteristics
of
obser-
vation
sequences
under
different
measurement
matrixs
are
analyzed,
and
row
echelon
matrix is
selected
as
the
measurement
matrix.
Then
wavelet decomposition with
the
observation
sequence
is
performed
and
low
frequency
coefficients
of
the
result
is
used
for
pitch
extraction through autocorrelation
method.
Simulation
and
analysis
show
that
compared
with
traditional
autocorrelation
method,
the
pitch
detection
accuracy
of
this
method
is
almost
the
same
as
that
of
the
traditional
method,
and
it
runs
well
when
the
speech
is
inter-
fused
with
Gaussian
white
noise,
but
the
calculation
amount
is
just
about
1/16
of
that
of
the
traditional
method.
Key
words:
speech
signal
processing;
pitch
period;
compressed
sensing
(CS)
;
non-reconstruction;
autocor-
relation
function;
wavelet decomposition
压缩感知（
Compressed
Sensing,
CS
）理论是
2004
年由
D.
Donoho
、
E.
Candes
等人提出的，它突破了传
统奈奎斯特采样定理对采样率与信号带宽之间的约
束关系，实现对信号的边采样边压缩，用更少的采样
数据量来恢复原始信号。二可信号在某个域满足稀疏
性，且感知矩阵满足
RIP
(
Restricted
Isom
et
町
Proper
ty
）特性时，信号能够被以大概率精确重构。文献［
1
-2
］指出语音信号在多个域满足稀疏性条件，若能
利用更少的信号量来提取原始信号的各种特征，将
大大减少计算量。本文主要研究将斥缩感知技术应
*
收稿日期：
2013
05
03
；修回日期：
2013
06
24
Received
date:
2013
05
03
;
Revised
date
:
2013
06
24
基金项目：国家自然科学基金资助项目（
61271335)
Foundation ltem·The
National
刊
atural
Science
卡。
undation
of
China
(No.
61271335)
'*
通讯作者：
jasthird@
163.
凹
m
Corresponding
author:jasthird@
163.
com
•
1327.
电讯技术
第
53
卷第
10
期
2013
年
10
月
Telecommunication Engineering
doi:
10.
3969/j.
issn.
1001-893x.
2013.
10.
014
Vol.
53
No.
10
Oct.
2013
基于压缩感知观测序列的语音信号基音周期提取＊
（南京邮电大学通信与信息工程学院，南京
210003)
摘
要：语音信号的清浊音辨别及基音周期的提耳又是语音信号处理的重要组成部分。提出了一种基
于压缩感知（
cs
）观测序列的语音基音周期才是取方法，大大减少了基音周期才是取的计算量。首先分
析不同观测矩阵下观测
l
序列的特征，选择了行阶梯矩阵作为所提方法的观测矩阵；然后对此观测矩
阵下的观测序列做小波分解，对小波分解的低频系数采用自相关方法提取基音周期。仿真与分析表
明，年口传统方法比较，所提方法与传统方法基音检测准确度相当，且在语音含有高斯白噪声的条件下
运行良好，但计算量仅约为传统方法的
1/160
关键词：语音信号处理；基音周期；压缩感知；非重构；自相关函数；小波变换
中图分类号：
TN912.
3
文献标志码：
A
文章编号：
1001-893X
(
2013)
10-1327-04
Pitch Detection Based on Compressed Sensing Sequence
CAO
Yang
(College
of
Telecommunication & lnfonnation Engineering,
Nanjing
University
of
Posts
and
Telecomrnunicalions,
Nanjing
2 l 0003, China)
Abstract:
Voiced/
unvoiced
decision
and
pitch
detection
technique
is
an
important
part
of
speech
signal
processing.
Based
on
the
observation
sequence
。
f
compressed
sensi
吨（
CS)
, a
method
to
detect
the
pitch
is
proposed
in
this
paper,
and
the
caculation
amount
is greatly
reduced
.
Firstly,
the
characteristics
of
obser-
vation
sequences
under
different
measurement
matrixs
are
analyzed,
and
row
echelon
matrix is
selected
as
the
measurement
matrix.
Then
wavelet decomposition with
the
observation
sequence
is
performed
and
low
frequency
coefficients
of
the
result
is
used
for
pitch
extraction through autocorrelation
method.
Simulation
and
analysis
show
that
compared
with
traditional
autocorrelation
method,
the
pitch
detection
accuracy
of
this
method
is
almost
the
same
as
that
of
the
traditional
method,
and
it
runs
well
when
the
speech
is
inter-
fused
with
Gaussian
white
noise,
but
the
calculation
amount
is
just
about
1/16
of
that
of
the
traditional
method.
Key
words:
speech
signal
processing;
pitch
period;
compressed
sensing
(CS)
;
non-reconstruction;
autocor-
relation
function;
wavelet decomposition
压缩感知（
Compressed
Sensing,
CS
）理论是
2004
年由
D.
Donoho
、
E.
Candes
等人提出的，它突破了传
统奈奎斯特采样定理对采样率与信号带宽之间的约
束关系，实现对信号的边采样边压缩，用更少的采样
数据量来恢复原始信号。二可信号在某个域满足稀疏
性，且感知矩阵满足
RIP
(
Restricted
Isom
et
町
Proper
ty
）特性时，信号能够被以大概率精确重构。文献［
1
-2
］指出语音信号在多个域满足稀疏性条件，若能
利用更少的信号量来提取原始信号的各种特征，将
大大减少计算量。本文主要研究将斥缩感知技术应
*
收稿日期：
2013
05
03
；修回日期：
2013
06
24
Received
date:
2013
05
03
;
Revised
date
:
2013
06
24
基金项目：国家自然科学基金资助项目（
61271335)
Foundation ltem·The
National
刊
atural
Science
卡。
undation
of
China
(No.
61271335)
'*
通讯作者：
jasthird@
163.
凹
m
Corresponding
author:jasthird@
163.
com
•
1327.
资源评论
