a sub-band filter design approach for sound field reproduction
TRANSCRIPT
Purdue University Purdue University
Purdue e-Pubs Purdue e-Pubs
Publications of the Ray W. Herrick Laboratories School of Mechanical Engineering
12-11-2020
A Sub-band Filter Design Approach for Sound Field Reproduction A Sub-band Filter Design Approach for Sound Field Reproduction
Yongjie Zhuang Purdue University, [email protected]
Guochenhao Song Purdue University, [email protected]
Yangfan Liu Purdue University, [email protected]
Follow this and additional works at: https://docs.lib.purdue.edu/herrick
Zhuang, Yongjie; Song, Guochenhao; and Liu, Yangfan, "A Sub-band Filter Design Approach for Sound Field Reproduction" (2020). Publications of the Ray W. Herrick Laboratories. Paper 225. https://docs.lib.purdue.edu/herrick/225
This document has been made available through Purdue e-Pubs, a service of the Purdue University Libraries. Please contact [email protected] for additional information.
PURDUE UN IVERS IT Y @
A Sub-band Filter Design Approach for Sound Field Reproduction
Yongjie Zhuang Guochenhao Song
Yangfan Liu
Ray W. Herrick Laboratories, school of Mechanical Engineering,
Purdue University [email protected]
PURDUE ~ RAY W. HERRICK1111 pi~
UNIVERSITY . LABORATORIES
Content โข Introduction
โข Methodology
โข Results
โข Conclusions
PURDUE ~ RAY W. HERRICK1111 pi~
UNIVERSITY . LABORATORIES
Content โข Introduction
โข Methodology
โข Results
โข Conclusions
PURDUE UNIVERSITY .
....
....
l
....
\
~ RAY W. HERRICK1111 pi~ LABORATORIES
Statement of research problem
Sound field reproduction uses loudspeakers to produce desired sound at locations.
When designing filter for sound that spans a wide frequency range:
Low frequency band longer time span Large number of filter coefficients High frequency band higher sampling frequency
Impulse response
๐๐
long time span
high sampling frequency
PURDUE ~ RAY W. HERRICK1111 pi~
UNIVERSITY . LABORATORIES
Statement of research problem
An approach is proposed to design filter in a sub-band form:
Design all sub-band filters directly in one optimization problem:
The transition region between two sub-band filters can be designed conveniently
The computational load can be reduced even if sub-band filters structure is not required
PURDUE ~ RAY W. HERRICK1111 pi~
UNIVERSITY . LABORATORIES
Content โข Introduction
โข Methodology
โข Results
โข Conclusions
,Ir
PURDUE ~ RAY W. HERRICK1111 pi~
UNIVERSITY . LABORATORIES
Designing filter directly
๐๐๐๐ desired
added delay signal d(n)
ฮฃy(n) +
e(n) -r(n) ๐พ๐พ๐๐ x(n)
๐ฏ๐ฏ output error filtered original
filter signal system signal signal signal response
Example: Cost function: Constraints: Use loudspeaker to produce Minimizing the power of Filter response ๐พ๐พ๐ฅ๐ฅ(f) desired sound at certain locations error signal ๐๐
-.
;-- - - - - - - - - - - -I
I I
I I I
I I
I ,Ir
I ,. I , . - - -
I - - - - . .
I ' ~ I
I I - I I
I I I I
I , ___________ _
PURDUE ~ RAY W. HERRICK1111 pi~
UNIVERSITY . LABORATORIES
Designing filter when sub-band technique is used
๐๐๐๐ desired
added delay signal d(n)
x(n) ฮฃ
๐พ๐พ1
sub-band filter 1
๐พ๐พ๐๐
sub-band filter N
+... y(n)
+ e(n) -r(n) ๐ฏ๐ฏ
output error filtered original signal system signal signal
signal response
๐พ๐พ๐๐
Example: Cost function: Constraints: Use loudspeaker to produce Minimizing the power of Filter response ๐พ๐พ๐๐ (f) desired sound at certain locations error signal ๐๐
( ) ( )~ ( ) [ -]
[
PURDUE ~ RAY W. HERRICK1111 pi~
UNIVERSITY . LABORATORIES
Expressing sub-band filters as one equivalent filter
Conventional method (one single filter)
frequency response of designed filter at frequency ๐๐๐๐ :
๐๐โ๐๐๐๐๐๐๐๐๐
๐๐โ๐๐๐๐๐๐๐๐๐(๐๐๐ก๐กโ1)= ๐น๐น ๐๐๐๐ , ๐๐๐ ๐ , ๐๐๐ก๐ก ๐ค๐ค๐ฅ๐ฅ , ๐น๐น ๐๐๐๐ , ๐๐๐ ๐ , ๐๐๐ก๐ก =๐๐๐ฅ๐ฅ ๐๐๐๐ 1 ๐๐๐ ๐ โฏ ๐๐๐ ๐
๐๐๐ ๐ is the sampling frequency, ๐๐๐ก๐ก is the number of filter coefficients, ๐ค๐ค๐ฅ๐ฅ is the filter coefficients
Sub-band structure
frequency response of designed filter at frequency ๐๐๐๐:
๐๐ ๐๐
โ๐๐๐๐๐๐๐๐๐ โ๐๐๐๐๐๐๐๐๐(๐๐๐ก๐กโ1) ๐๐๐๐ (๐๐๐๐) = ๐๐๐ ๐ ๐๐ ๐๐๐ ๐ ๐๐ ๐ค๐ค๐๐1 ๐๐ โฏ ๐๐
๐๐=1 ๐๐=1
( )
( ) [ ( ) --+
( ) [ ( ) ( )]
PURDUE ~ RAY W. HERRICK1111 pi~
UNIVERSITY . LABORATORIES
Expressing sub-band filters as one equivalent filter
So designing sub-band filters can be treated as: designing one filter ๐ค๐ค๐ฅ๐ฅ with modified Fourier matrix ๐น๐น ๐๐๐๐
๐๐ ๐๐
๐๐โ๐๐๐๐๐๐๐๐๐
๐๐โ๐๐๐๐๐๐๐๐๐(๐๐๐ก๐กโ1)
๐๐๐ฅ๐ฅ ๐๐๐๐ = ๐๐๐๐ (๐๐๐๐) = ๐๐๐ ๐ ๐๐ ๐๐๐ ๐ ๐๐ ๐ค๐ค๐๐ = ๐น๐น ๐๐๐๐ ๐ค๐ค๐ฅ๐ฅ ,1 โฏ๐๐=1 ๐๐=1
๐น๐น ๐๐๐๐ = ๐น๐น ๐๐๐๐ , ๐๐๐ ๐ 1 , ๐๐๐ก๐ก1
โฏ ๐น๐น ๐๐๐๐ , ๐๐๐ ๐ ๐๐ , ๐๐๐ก๐ก๐๐
,
๐ค๐ค1๐ค๐ค๐ฅ๐ฅ = โฎ
๐ค๐ค๐๐
So all the sub-band filters can be designed in one optimization problem if designed in the frequency domain. The transition region can be designed more conveniently.
....
...
..... ..... .... ... ... ....
... ยทยทยท ยทยท ... .. . . . ยทยทยทยทยทยทยทยท .. ... . . . . ยท .ยทยทยท .. ... ยทยท .. ยทยท ...
(\~~ :~)') ยทยทยทยทยทโขโขยท ..
ยทยทยท ....
ยทยทยทยทยทยทยทยทยทยทยทยทยทยทยทยทยทยทยทยทยทยทยทยท ยทยทยท ...
PURDUE ~ RAY W. HERRICK1111 pi~
UNIVERSITY . LABORATORIES
Overview of proposed design process
Conventional method General Nonlinear algorithm
(e.g., sequential quadratic programming)
Design Problem
Convex Optimization
Proposed method Primal-dual Cone
Programming Interior-point Global minimum Possible to apply โข Efficient Cost function: usually guaranteed efficient algorithms โข Upper-bound for
iteration times Power of ๐๐
Constraints: Filter response
I C )I .... - ( ( ( )
' ( )
I C )I ( ) ---+ .. PURDUE ~
RAY W. HERRICK1111 pi~ UNIVERSITY . LABORATORIES
Problem formulation
Design Problem Expressed in Convex Problem Cost function: Total power of e:
๐๐2 ๐๐2 ๐๐2 ๐๐2๐ ๐ค๐ค๐ฅ๐ฅ
T ๐ด๐ด๐ฝ๐ฝ ๐๐๐๐ ๐ค๐ค๐ฅ๐ฅ + 2Re ๐๐๐ฝ๐ฝT ๐๐๐๐ ๐ค๐ค๐ฅ๐ฅ + ๐๐๐ฝ๐ฝ ๐๐๐๐,๐ธ๐ธ ๐๐๐๐ ๐๐=๐๐1 ๐๐=๐๐1 ๐๐=๐๐1
๐๐=๐๐1 Convex โข Quadratic
Constraints: โข ๐ด๐ด๐ฝ๐ฝ ๐๐๐๐ p.s.d
Filter response: The magnitude of frequency response:
Vector norm Convex โ๐น๐น ๐๐๐๐ , ๐๐๐ ๐ 1 , ๐๐๐ก๐ก1
๐ค๐ค๐๐โ๐ โ ๐ถ๐ถ๐๐ (๐๐๐๐) โค 0โค ๐ถ๐ถ๐๐(๐๐๐๐)๐๐๐๐ ๐๐๐๐
( ( )
? โข ( ) -+ ....
PURDUE ~ RAY W. HERRICK1111 pi~
UNIVERSITY . LABORATORIES
Cone Programming Reformulation
Convex Problem Standard Cone Programming Cost function:
Cost function: ๐๐T๐ฅ๐ฅ ๐๐2 ๐๐2
๐ค๐ค๐ฅ๐ฅT
๐ด๐ด๐ฝ๐ฝ ๐๐๐๐ ๐ค๐ค๐ฅ๐ฅ + 2Re ๐๐๐ฝ๐ฝT ๐๐๐๐ ๐ค๐ค๐ฅ๐ฅ + ๐๐๐ฝ๐ฝ ๐๐๐๐
๐๐2
๐๐=๐๐1 ๐๐=๐๐1 ๐๐=๐๐1 Constraints: ๐ฅ๐ฅ โ ๐พ๐พ๐๐ , ๐๐ = 1, 2, 3 โฆ
Constraints: ๐ด๐ด๐ฅ๐ฅ = ๐๐
๐๐ to be a constant vector โ๐น๐น ๐๐๐๐ , ๐๐๐ ๐ 1 , ๐๐๐ก๐ก1
๐ค๐ค๐๐โ๐ โ ๐ถ๐ถ๐๐(๐๐๐๐) โค 0 ๐พ๐พ๐๐ to be a convex cone
๐ด๐ด, ๐๐ to be a constant matrix and vector
..
PURDUE UNIVERSITY .
.. .. ..
..
.. ~ RAY W. HERRICK1111 pi~ LABORATORIES
Cone Programming Reformulation
Convex Problem Cone Programming โข Reformulate quadratic cost function
Cost function: ๐ฅ๐ฅT ๐ด๐ด ๐ฅ๐ฅ + ๐๐T๐ฅ๐ฅ + ๐๐
Cost function:
Constraints: ๏ฟฝฬ๏ฟฝ๐ก0 = 1
Linear cost function
Rotated second-order cone
Linear constraint
โข The vector norm constraint
Constraints: โ๐ฅ๐ฅโ๐ โ ๐๐ โค 0
โ ๐ด๐ด ๐ฅ๐ฅโ๐ โค ๐ก๐ก0 ๏ฟฝฬ๏ฟฝ๐ก0
๐ก๐ก0 + ๐๐T๐ฅ๐ฅ
Constraints: โ๐ฅ๐ฅโ๐ โค ๐ก๐ก Second-order cone
๐ก๐ก = ๐๐ Linear constraint
( ( ) ( ) --+
( ) -+ (
PURDUE ~ RAY W. HERRICK1111 pi~
UNIVERSITY . LABORATORIES
Cone Programming Reformulation
Convex Problem Cone Programming
Cost function: Cost function: ๐๐2๐๐2 ๐๐2 ๐๐2T
๐ค๐ค๐ฅ๐ฅ ๐ด๐ด๐ฝ๐ฝ ๐๐๐๐ ๐ค๐ค๐ฅ๐ฅ + 2Re ๐๐๐ฝ๐ฝT ๐๐๐๐ ๐ค๐ค๐ฅ๐ฅ + ๐๐๐ฝ๐ฝ ๐๐๐๐ ๐ก๐ก0 + 2Re ๐๐๐ฝ๐ฝT ๐๐๐๐ ๐ค๐ค๐ฅ๐ฅ๐๐=๐๐1 ๐๐=๐๐1 ๐๐=๐๐1 ๐๐=๐๐1
Constraints: Constraints:
โ๐น๐น ๐๐๐๐ , ๐๐๐ ๐ 1 , ๐๐๐ก๐ก1
๐ค๐ค๐๐โ๐ โ ๐ถ๐ถ๐๐(๐๐๐๐) โค 0 โ๐น๐น ๐๐๐๐ , ๐๐๐ ๐ 1 , ๐๐๐ก๐ก1 ๐ค๐ค๐๐โ๐ โค ๐ก๐ก3,๐๐ ,
๐ก๐ก3,๐๐ = ๐ถ๐ถ(๐๐๐๐)
( )
( )
/ 1 ' ()
l
( ) [ -] PURDUE ~
RAY W. HERRICK1111 pi~ UNIVERSITY . LABORATORIES
A reduced order technique
Sometimes, the designed filter has high frequency response concentrated in small time span: ๐๐๐๐ ๐๐ Number of
coefficients: ๐๐๐๐ ๐๐
๐๐ ๐๐ ร 0.1๐๐๐ ๐
๐๐ 0.2๐๐ ร ๐๐๐ ๐ ๐๐๐๐ ๐๐
๐๐ 0.1๐๐ ร ๐๐๐ ๐ Number of coefficients: ๐๐ ร ๐๐๐ ๐ ๐๐
In this case, ๐ค๐ค๐๐ (with higher sampling frequency) can be chosen to start with ๐ก๐ก = ๐๐ฮ , where ๐๐ > 0 , then we have:
๐๐โ๐๐๐๐๐๐๐๐๐๐๐
๐๐โ๐๐๐๐๐๐๐๐๐(๐๐+1)
๐๐โ๐๐๐๐๐๐๐๐๐(๐๐๐ก๐กโ1)๐น๐น๐๐ ๐๐๐๐ , ๐๐๐ ๐ , ๐๐๐ก๐ก = ๐๐๐ ๐ ๐๐๐ ๐ โฏ ๐๐๐ ๐
PURDUE ~ RAY W. HERRICK1111 pi~
UNIVERSITY . LABORATORIES
Content โข Introduction
โข Methodology
โข Results
โข Conclusions
/
[)J:
PURDUE ~ tv\.'I'. VY. TTCKJV\......1'.. ... ~
UNIVERSITY . LABORATORIES
Experimental setup โข An experimental setup for psychoacoustic listening test โข Speaker should produce desired sound at listening location
Listening location Speaker
PURDUE ~ RAY W. HERRICK1111 pi~
UNIVERSITY . LABORATORIES
Experimental setup Required sampling frequency: 48 kHz (๐ซ๐ซ =20.83 us)
Desired delay: 19200 ๐ซ๐ซ
Two sub-band filters:
Sampling frequency Filter coefficients Starting time
Filter 1 2.4 kHz 1920 0 Filter 2 48 kHz 3000 17700 ๐ซ๐ซ
SeDuMi is used to solve the reformulated cone programming problem
PURDUE ~ RAY W. HERRICK1111 pi~
UNIVERSITY . LABORATORIES
Result
-40
-20
0
20
ampl
itude
(dB)
w1 w2
The frequency response of both filter around 1200 Hz
0 500 1000 1500 2000
frequency (HZ)
( ) ( ) I I I I I
I I I I
I I I I I
X
X
X
PURDUE ~ RAY W. HERRICK1111 pi~
UNIVERSITY . LABORATORIES
0
Result
The frequency response of H(f)๐๐๐ฅ๐ฅ ๐๐ H(f) ๐๐๐ฅ๐ฅ ๐๐ around 1200 Hz 5
ampl
itude
(dB) 0
ampl
itude
(dB)
-20
-5-40
0 0.5 1 1.5 2 -10 600 800 1000 1200 1400 1600 1800
10 4frequency (HZ) frequency (HZ)
10 4
0 -2000
-2
phas
e (ra
d)
phas
e (ra
d)
-4
-6
-3000
-40000 0.5 1 1.5 2
frequency (HZ) 10 4 600 800 1000 1200 1400 1600 1800
frequency (HZ)
PURDUE ~ RAY W. HERRICK1111 pi~
UNIVERSITY . LABORATORIES
Result Combining two sub-band filters together in time domain
1500 The combination is done by:
1000
โข Upsampling the sub-band filter 1 500
0 0.2 0.4 0.6
time (s)
with lower sampling frequency
โข Adds the upsampled filter 1 with ampl
itude
0
filter 2 -500
-1000
-1500
The designed filter coefficients are 1920+3000 = 4920, which is much smaller than 48000 ร 19๐0
๐400 = 38400
PURDUE ~ RAY W. HERRICK1111 pi~
UNIVERSITY . LABORATORIES
Content โข Introduction
โข Methodology
โข Results
โข Conclusions
PURDUE ~ RAY W. HERRICK1111 pi~
UNIVERSITY . LABORATORIES
Conclusions
The proposed method can design sub-band filters for sound field reconstruction in one optimization problem, so designing transition region is more convenient.
The optimization problem can be reformulated to a convex problem, then further reformulated to a cone programming problem. These guarantees the global optimal solution can be found in an efficient way.
A reduced-order technique can be used to reduce the variables in filter design problem if different frequency bands of required filter have impulse response concentrated in different time intervals.