On Microphone On Microphone On Microphone On Microphone On Microphone On Microphone On Microphone
On Microphone---Array Beamforming Array Beamforming Array Beamforming Array Beamforming Array Beamforming Array Beamforming Array Beamforming Array Beamforming from a MIMO Acoustic Signal
from a MIMO Acoustic Signal from a MIMO Acoustic Signal from a MIMO Acoustic Signal from a MIMO Acoustic Signal from a MIMO Acoustic Signal from a MIMO Acoustic Signal from a MIMO Acoustic Signal
Processing Perspective Processing Perspective Processing Perspective Processing Perspective Processing Perspective Processing Perspective Processing Perspective Processing Perspective
Jacob Benesty, Jingdong Chen, Yiteng (Arden) Haung and Jacek Dmochowski _________________________________________________________________________________
__________________________________________________________________________________________________________________________________________________________________
_________________________________________________________________________________
Presented by
Osnat Goren-Peyser 24
24 24
24 June June June 2007 June 2007 2007 2007
Outline Outline Outline Outline Outline Outline Outline Outline
• Introduction
• Problem Description
• Beamforming
• Filer and sum technique
• Filer and sum technique
• Estimation algorithms
• Experimental results
• Further work
• Summery
Abbreviation Abbreviation Abbreviation Abbreviation Abbreviation Abbreviation Abbreviation Abbreviation
• BF – Beamforming
• DOA – Direction Of Arrival
• LS – Least Squares
• MN – Minimum Norm
• LCMV – Linearly Constrained Minimum Variance
• LCMV – Linearly Constrained Minimum Variance
• MVDR – Minimum Variance Distortionless Response
• GSC – Generalize Sidelobe Canceller
• SIR – Signal to Interference Ratio
Introduction Introduction Introduction Introduction Introduction Introduction Introduction Introduction
• Signal enhancement problem using microphone array.
• The objective of the array processing is to estimate the desired signals from the given microphones outputs.
• Noise and interferences:
– Noise
– Reverberation – the persistence of sound in a particular space after the original sound is removed.
– Reverberation – the persistence of sound in a particular space after the original sound is removed.
– Interferers – signals with the same frequency as the desired signal that do not arrive from the desired signal DOA.
• Microphone arrays consist of sets of microphone sensors that are spatially arranged in specific pattern
– Different microphones receive different signals
– Can be exploited to separate the desired signals from the interferers using a beamformer.
Problem Description Problem Description Problem Description Problem Description Problem Description Problem Description Problem Description Problem Description
• M sources
– P desired signals – Q interferers – M = P+Q
• N microphones
• We assume N≥M
• sm(k) - The mth source signal
s8(k)
x8(k)
x9(k)
h11
h21
h31
Sources Channel Microphones
• sm(k) - The mth source signal
• xn(k) - The output of the nth microphone
• hnm - The acoustic channel impulse response from source m to microphone n
• Lh – The length of the longest channel impulse response
• bn(k) - The noise observed at the nth microphone
s9(k)
x:(k)
sM(k)
xN(k)
hNM h31
hN1
Mathematical Model Mathematical Model Mathematical Model Mathematical Model Mathematical Model Mathematical Model Mathematical Model Mathematical Model
• The output of the nth microphone at time k:
• Assumptions:
• Assumptions:
– N≥M
– LTI system
– Discarding the noise terms: b(k)=0 for all k
– The first P signals are the desired sources, and the rest Q sources are interferers.
– Source signals in far-field
Interferences Interferences Interferences Interferences Interferences Interferences Interferences Interferences
• Reverberation
– Reverberation is the collection of all reflected sounds in a room.
– Reverberation time is the time required for a sound in a room to decay by 60 dB (called RT60)
– Self interferences – Proportional to Lh – Proportional to Lh
• Interferers
– signals with the same frequency as the desired signal that do not arrive from the desired signal DOA.
– External interferences – Proportional to Q
• Measurements
– Speech dereverberation – Interference suppression
The main goal The main goal The main goal The main goal The main goal The main goal The main goal The main goal
• We want to perfectly estimate the desired signals from the microphone observations using weights
• Find s.t .
••
• Constraints:
Beamforming Beamforming Beamforming Beamforming Beamforming Beamforming Beamforming Beamforming
• A Beamformer is an array of sensors which can do spatial filtering.
• The objective is to estimate the signal arriving from the desired direction in the presence of noise and other interfering signals.
• A beamformer does spatial filtering in the sense that it separates two signals with overlapping frequency content originating from different that it separates two signals with overlapping frequency content originating from different directions.
• Beamforming is applicable to either transmission or reception of energy.
– The paper discuss formation of beams for reception
• By weighting or filtering and then summing signals from different microphones, the beamformer can produce virtual directivity pattern (=weighted sum of individual patterns)
Beamformer Classification Beamformer Classification Beamformer Classification Beamformer Classification Beamformer Classification Beamformer Classification Beamformer Classification Beamformer Classification
• Beamformers are classified as either data independent or statistically optimum, depending on how the weights are chosen.
• The weights in a data independent
beamformer do not depend on the array data and are chosen to present a
specified response for all signal and specified response for all signal and interference scenarios.
• The weights in a statistically optimum beamformer are chosen based on the statistics of the array data to optimize the array response.
– The weighs can be either fixed or adaptively determined.
– The statistics of the array data are not
usually known and may change over time so adaptive algorithms are typically used.
Filter and sum Beamforming Filter and sum Beamforming Filter and sum Beamforming Filter and sum Beamforming Filter and sum Beamforming Filter and sum Beamforming Filter and sum Beamforming Filter and sum Beamforming
• Microphone signals are filtered and then summed together.
• {g
n,l} are the lth
coefficient of microphone n FIR filter
n FIR filter
• L = FIR filter length
– L=8 Delay and sum
• Filter coefficients and
filter length are chosen
depends on the desired
performance.
Filter and sum Beamforming Filter and sum Beamforming Filter and sum Beamforming Filter and sum Beamforming Filter and sum Beamforming Filter and sum Beamforming Filter and sum Beamforming Filter and sum Beamforming
Estimation problem Estimation problem
• In order to estimate the desired signals from the microphone observation , we should determine NP L
g-length FIR filters
• Find the optimal BF weights
– Determine the optimization criterion – Determine the optimization criterion
• Determine the BF algorithm
• Different optimization criterion Different performance !
– Determine the set of linear equations describing the estimation
• Full channel knowledge
• Partial channel knowledge
• Find bounds on FIR filter length
Filter and sum Beamforming Filter and sum Beamforming Filter and sum Beamforming Filter and sum Beamforming Filter and sum Beamforming Filter and sum Beamforming Filter and sum Beamforming Filter and sum Beamforming
Analysis Analysis
• The following linear set of equations describing the estimation problem:
H
Tg p =u
p’
Where:
– HT MLxNLgchannel matrix – gp NLg length column vector – gp NLg length column vector – up’ ML length column vector – L = Lg+Lh-8
• Channel knowledge:
– Full knowledge of the impulse response matrix H
– Partial Knowledge of the impulse responses from source p to the N microphones
Filter and sum Beamforming Filter and sum Beamforming Filter and sum Beamforming Filter and sum Beamforming Filter and sum Beamforming Filter and sum Beamforming Filter and sum Beamforming Filter and sum Beamforming
Analysis (cont.) Analysis (cont.)
• The following linear sub set of equations describing the estimation problem in the case that only the impulse responses from source p to the N microphones are known:
H
:pTg
p=u
pWhere:
– H:pT LxNLgchannel matrix – gp NLg length column vector – up L length column vector – up L length column vector – L = Lg+Lh-8
• The following linear sub set of equations describing the estimation problem in the case that only the impulse responses from source p to the N microphones are known:
h
:pT(k
p)g
p=8
Where:
– h:p(kp) the kpth column of H:p
– h:p(kp)T NLglength channel vector – gp NLg length column vector
Deterministic characterization
Deterministic characterization
Deterministic characterization
Deterministic characterization
Deterministic characterization
Deterministic characterization
Deterministic characterization
Deterministic characterization
of estimation problem I of estimation problem I of estimation problem I of estimation problem I of estimation problem I of estimation problem I of estimation problem I of estimation problem I
Deterministic characterization Deterministic characterization Deterministic characterization Deterministic characterization Deterministic characterization Deterministic characterization Deterministic characterization Deterministic characterization of estimation problem II
of estimation problem II
of estimation problem II
of estimation problem II
of estimation problem II
of estimation problem II
of estimation problem II
of estimation problem II
Deterministic characterization Deterministic characterization Deterministic characterization Deterministic characterization Deterministic characterization Deterministic characterization Deterministic characterization Deterministic characterization of estimation problem III
of estimation problem III
of estimation problem III
of estimation problem III
of estimation problem III
of estimation problem III
of estimation problem III
of estimation problem III
Filter and sum Beamforming Filter and sum Beamforming Filter and sum Beamforming Filter and sum Beamforming Filter and sum Beamforming Filter and sum Beamforming Filter and sum Beamforming Filter and sum Beamforming
Stochastic characterization Stochastic characterization Stochastic characterization Stochastic characterization Stochastic characterization Stochastic characterization Stochastic characterization Stochastic characterization
• R
ssis the sources covariance matrix
– R
ssis MLxML matrix – R
sshas full rank
– H
Thas full column rank – H has full column rank
• R
xxis the microphones covariance matrix
– R
xx=HR
ssH
T– R
xxis NL
gxNL
g– R
xxis invertible
Beamforming Algorithms
Beamforming Algorithms Beamforming Algorithms
Beamforming Algorithms Beamforming Algorithms
Beamforming Algorithms Beamforming Algorithms
Beamforming Algorithms
Known solutions for a system Known solutions for a system Known solutions for a system Known solutions for a system Known solutions for a system Known solutions for a system Known solutions for a system Known solutions for a system of linear equations
of linear equations of linear equations of linear equations of linear equations of linear equations of linear equations of linear equations
• We want to find x for the equation Ax=b
– A is a known MxN matrix (usually M>N) – x is an unknown Nx8 vector
– b is a known Mx8 vector
• If A is a square matrix (M=N) with full rank the
• If A is a square matrix (M=N) with full rank the solution x=A
-8b is unique.
• For M>N
– Ax≈b
– Least Squares (LS) solution
– Minimum norm (MN) solution
Known solutions for a system Known solutions for a system Known solutions for a system Known solutions for a system Known solutions for a system Known solutions for a system Known solutions for a system Known solutions for a system of linear equations (cont.)
of linear equations (cont.) of linear equations (cont.) of linear equations (cont.) of linear equations (cont.) of linear equations (cont.) of linear equations (cont.) of linear equations (cont.)
LS LS LS LS LS LS LS LS
• Minimizing the Euclidean norm squared of Ax-b.
• Optimization criterion:
MN MN MN MN MN MN MN MN
• Minimizing the L9 norm of x
• Optimization criterion:
• Solution:
• Assumptions:
– M>N
– A has full column rank : rank (A)=N
• Solution:
• Assumptions:
– M>N
– A has full column rank : rank (A)=N
LS LS LS LS LS LS LS LS
• LS=Least Squares
• Estimation problem: H
Tg
p=u
p’
• Minimizing the Euclidian distance of the residual error: H
Tg
p-u
p’
• The optimization problem:
• The solution is: g
p=(HH
T)
-8Hu
p’
• Bound on L
g – M=N• No upper bound on Lg
– N>M
• rank(HT)=NLg
•
LS (cont.) LS (cont.) LS (cont.) LS (cont.) LS (cont.) LS (cont.) LS (cont.) LS (cont.)
• Special case: N>M s.t and integer
– H
Tis a square matrix and g
p=(H
T)
-8u
p’
– Can perfectly estimated the source signals!
– It is better to have more microphones than sources!
• Advantage
– Data independent BF
• Disadvantage
– Not flexible – the whole impulse response matrix H must
be known.
LCMV Filter LCMV Filter LCMV Filter LCMV Filter LCMV Filter LCMV Filter LCMV Filter LCMV Filter
• LCMV=Linearly Constrained Minimum Variance
• Minimizing the BF output variance (power) while maintaining m linear constraints
– Forcing signals from the direction of interest to pass with specific gain and phase – Forcing zero gain in interferer’s direction
• Each linear constraint uses one degree of freedom in the weight vector Only n-m degree of freedom available for minimizing output variance.
• Definitions:
• Definitions:
– A is mxn constraints matrix – b is mx8 desired response vector
• Assumptions:
– m linear independent constraints rank(A)=m – m<n
• Problem:
• Solution (based on Lagrange multipliers method):
LCMV LCMV LCMV LCMV LCMV LCMV LCMV LCMV
• Advantage
– Flexible – forming the beam using general constraints.
• Disadvantage
• Disadvantage
– Complexity – Computation of constrained weigh vector.
• Estimation problem:
– LCMV8: Partial channel knowledge and N≥M
– LCMV9: Full channel knowledge and N>M
LCMV LCMV LCMV LCMV LCMV LCMV LCMV LCMV1 1 1 1 1 1 1 1
• Estimation problem: H:pTgp=up
• The optimization problem:
• The solution is:
• Bounds on Lg:
– N=M
••
• No upper bound
– N>M
•
• Special case: N=M s.t and integer H:p is square matrix and gp=(H:pT)-8up
LCMV LCMV LCMV LCMV LCMV LCMV LCMV LCMV2 2 2 2 2 2 2 2
• Estimation problem: H
Tg
p=u
p’
• The optimization problem:
• The solution is:
• The solution is:
• Bound on Lg:
• Special case: for and integer
– H is square matrix and g =(H
T)
-8u ’
GSC GSC GSC GSC GSC GSC GSC GSC
• GSC=Generalized Sidelobe Canceller
• Transforms the LCMV 8 algorithm from a constrained problem into an unconstrained form.
• gp is divided into two components operating on orthogonal subspaces:
gp=fp-Bpwp – Where:
– Where:
– fpis the MN solution of H:pTfp=up fp=H:p(H:pTH:p)-8up
– Bp is blocking matrix that spans the null space of H:p H:pT Bp=0
• The unconstraint optimization problem:
• The solution is:
– fpis a data independent BF (MN BF) – wpis a statistically optimum BF
• It can be shown that: gpLCMV8=gpGSC
GSC (cont.) GSC (cont.) GSC (cont.) GSC (cont.) GSC (cont.) GSC (cont.) GSC (cont.) GSC (cont.)
• LCMV8 is the sum of two orthogonal vectors: fp and Bpwp
– The objective of fp is to perform dereverberation – fixed BF is sufficient to perform dereverberation – The objective of -Bpwp is to reduce interference
• Increasing Lg:
– will not change the dereverberation performance.
– will not change the dereverberation performance.
– A better interference suppression is expected.
• As reverberation of the room increases (Lh↑), interference suppression decreases.
– Perfect dereverberation is possible (if H:p can be accurately estimated) BUT perfect interference suppression is not!
– One way for improvement:
LCMV filter for dereverberation followed by Wiener filter for noise reduction.
• The dimension of the nullspace of H:pT=NL -L
GSC (cont.) GSC (cont.) GSC (cont.) GSC (cont.) GSC (cont.) GSC (cont.) GSC (cont.) GSC (cont.)
• Bounds on L
wp– M=N
• No upper bound on Lg Lwpcan be as large as we wish
• Lg↑ better interference suppression
– M>N
• Lwp=(N-8)Lg-Lh-8
• Lwp=(N-8)Lg-Lh-8
• Based on the bound on Lg:
• There is a limit to interference suppression!!
• Special case: M=Q+8=N-K,K>0
– The upper bound of Lwp depends on: N,Q,Lh
– Q,N are fixed: Lg↑ we have to use a larger Lwp – Lh,N are fixed: Q↑ we have to use a larger Lwp
GSC (cont.) GSC (cont.) GSC (cont.) GSC (cont.) GSC (cont.) GSC (cont.) GSC (cont.) GSC (cont.)
• Disadvantage:
– Can be done only for LCMV8
– As L
hand Q increases the GSC problem (L
wp) becomes more difficult to solve we should expect performance
degradation related to LCMV8 performance.
degradation related to LCMV8 performance.
• Advantages:
– Easy for analysis
– Helps to understand the LCMV8 problem
– Less complicated implementation than LCMV8 BF.
MVDR MVDR MVDR MVDR MVDR MVDR MVDR MVDR
• MVDR=Minimum Variance Distortionless Response
• A special case of LCMV8 where we maintaining only the desired signal gain.
• The constraint: h
:pT(k
p)g
p=8
• The constraint: h
:pT(k
p)g
p=8
• The aim of the constraint is to align the desired source at the BF output
• The optimization problem:
• The solution is:
MVDR (cont.) MVDR (cont.) MVDR (cont.) MVDR (cont.) MVDR (cont.) MVDR (cont.) MVDR (cont.) MVDR (cont.)
• Bound on L
g: L
g≥k
p– Special case: L
g=k
pclassical delay and sum BF
• Advantages:
– Most useful in practice since it only requires the – Most useful in practice since it only requires the
knowledge of the relative delays among microphones
• Disadvantages:
– Not the best performance!
Experiments Experiments Experiments Experiments Experiments Experiments Experiments Experiments
• Goals
– Studying the effect of filter length on beamforming performance
– Comparing the different algorithms via simulations in realistic acoustic environments
• Setup
– Rectangular Room 6.7m long by 6.8m wide by 9.9m high
– Linear array of 4 omni-directional microphones
microphones – M=: (P=8,Q=9) – N=4
– s8 – desired source(mail speaker)
– s9,s: – interferences from the same female speaker.
– Lh was truncated
– Reverberation time: T60=0.:5s
• Experiment 1 – a priori knowledge
• Experiment 2 – no knowledge (using blind technique)
Experimental Results Experimental Results Experimental Results Experimental Results Experimental Results Experimental Results Experimental Results Experimental Results
Beamformer output Desired Signal & output of microphone 1
S1(k) y(k)
x1(k)
M=:
N=4 K=8 Lh=64 Lg=889 Lg=889
Measurement Measurement Measurement Measurement Measurement Measurement Measurement
Measurement Tools Tools Tools Tools Tools Tools Tools Tools
• SIR gain in dB
• SIR[dB]=SIRout[dB]-SIRin[dB]
• Where:
• IS
• Performs a comparison of the
spectral envelope between the clean and the desired signal
• IS 0 means perfectly dereverberation
• SIR gain > 800dB means perfectly interference suppression
Experimental Results Experimental Results Experimental Results Experimental Results Experimental Results Experimental Results Experimental Results Experimental Results
A –priori knowledgeA knowledgelind nique
Conclusions Conclusions Conclusions Conclusions Conclusions Conclusions Conclusions Conclusions
• LS and LCMV9 performance are the same
• As Lh increases (with the maximum Lg) SIR gain decreases
• When Lg is set to its maximum value both the LS and LCMV9 can achieve almost perfect interference suppression and speech dereverberation (SIR gain > 800dB & IS 0)
• LCMV8 and GSC can perform perfect speech dereverberation (IS 0 for all Lg,Lh) BUT their interference suppression performance is limited
all Lg,Lh) BUT their interference suppression performance is limited
• For a fixed Lh, reducing Lg decreases the amount of interference suppression significantly for all the methods, except MVDR.
• For all algorithms IS 0 indicating that these techniques have accomplished good speech dereverberation
• MVDR methods is very robust to the changes of both Lh and Lg and when Lg is small this method can achieve more interference suppuration then the other methods BUT the IS measures are very large
Conclusions (cont.) Conclusions (cont.) Conclusions (cont.) Conclusions (cont.) Conclusions (cont.) Conclusions (cont.) Conclusions (cont.) Conclusions (cont.)
• When all the techniques suffer from some but not significant performance degradation
• When
– LS and LCMV9 suffer significantly performance – LS and LCMV9 suffer significantly performance
degradation
– LCMV8 and GSC suffer some but not serious performance degradation
• MVDR performance does not deteriorate much as
L
hdecreases (comparing to case)
Further Work Further Work Further Work Further Work Further Work Further Work Further Work Further Work
• Option I – Resistance of GSC algorithm to channel dynamic.
• Option II – Comparison between the far-field GSC performance and the near-field GSC
GSC performance and the near-field GSC
algorithm.
Option I Option I Option I Option I Option I Option I Option I Option I
• Define a model for channel matrix changes
• Comparison between the GSC performance with and without channel dynamic.
• Based on Israel Cohen work: “Analysis of Two-
• Based on Israel Cohen work: “Analysis of Two-
Channel Generalized Sidelobe Canceller(GSC)
with Post-Filtering”.
Option I (cont.) Option I (cont.) Option I (cont.) Option I (cont.) Option I (cont.) Option I (cont.) Option I (cont.) Option I (cont.)
• Reference case – fixed s8,s9
– M=9 (P=8,Q=8) – N=:
• Case study 8 – s8 was moved, fixed s9
• Case study 9 – s8,s9 were moved
Option II Option II Option II Option II Option II Option II Option II Option II
• Define the near-field GSC problem
• Comparison between the far-filed and the near-field GSC algorithm performance.
• Based on Iain A. McCowan, Darren C. Moore
• Based on Iain A. McCowan, Darren C. Moore
ans S. Sridharan work: “Near-field Adaptive
Beamformer for Robust Speech Recognition”
םוכיס םוכיס –
ת – ת הנכה תולאשל תובוש הנכה תולאשל תובוש
• היעבה דה –
ע שער תשג
"
םינופורקימ ךרעמב שומיש י
• הרטמ עש –
ע רוקמ תוא ךור
"
םיעירפמ יוכידו םידה לוטיב י
• היעבה ןורתפל תומייק תושיג רואית תוקינכט –
המולא בוציע (Beamforming)
• רמאמב עצומה ןורתפה תגצה תאוושה –
תושיגב םידה לוטיבו םיעירפמ יוכיד ןבומב םיעוציב
beamforming םינוש הערפה יראתמ תחת תונוש
• השיג לכב שמתשהל יאדכ יתמ
?
– לש םלהל הבוגתה ךרוא רשאכ עודי ץורעה
ב שומישל תופידע הנשי ו LS
LCMV9
ה ןנסמשכ LCMV9 ו LS ב שומישל תופידע הנשי עודי ץורעה לש םלהל הבוגתה ךרוא רשאכ – ה ןנסמשכ
beamformer וכרעב
ילמיסקמה
– לש םלהל הבוגתה ךרוא רשאכ עודי וניא ץורעה
ב שומישל תופידע הנשי LCMV8
ו GSC
ה ןנסמשכ beamformer
וכרעב ילמיסקמה
– רתוי הברה MVDR
יטסובור רצק ןנסמ תריחבלו ץורע ךורעש תואיגשל
, לוטיב תלוכי םלוא
ידמל השלח ולש םידהה ,
הדובע יאנתב םג םיילמיטפוא
) רצק Lh , ךורעשו ךורא Lg קייודמ Lh
.(
רתויב ןטקה אוה ולש םיצוליאה טסש איה ךכל הביסה ,
שומימל רתוי לק םג אוה ןכל .
יאדכו
םידה לוטיבל תפסונ הקינכט םע בולישב קר וב שמתשהל .
• תירשפא הבחרה תגצה