Microphone Array Processing

• Estimation of the ATF

• Estimation of Rs
• GEVD and EVD

Richard C. Hendriks
May 30, 2023 1
Summary of Previous Lecture
 Microphone Measurement Model

Single microphone model:

<latexit sha1_base64="+GQMWBaldrmP5BbWJyU8JaDJLJY=">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</latexit>

d
X
x[n] = (hi ⇤ si )[n] + n[n]
i=1

Assumptions: Sources are assumed to be

• Additive
• zero-mean and mutually uncorrelated, i.e., E[si ] = 0, E[n] =
0, E[si sj ] = 0 8i, j and E[si n] = 0 8i.
• short-time stationary.
Validity of these assumptions?
 Short-Time Frequency Transform
Processing is often done in the so-called short-time frequency domain, i.e.,
<latexit sha1_base64="PEnwdiAXwk/RDvAJ5La+dt4hBZM=">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</latexit>

FFT on short windowed time frames.

• Time frames should obey Short time WSS assumption.
• STFT makes convolutive model (aproximately) multiplicative AND
helps to satisfy narrowband assumption.
Pd
• x(k, l) = i=1 ai (k, l)si (k, l) + n(k, l)

• For M microphones using stacked vector notation:

d
X
x(k, l) = ai (k, l)si (k, l) + n(k, l)
i=1

• Notice: As all processing is often done independently per frequency

band and time frame, time and frequency indices are usually neglected.
 Problem formulation
<latexit sha1_base64="XubE41DpJIFFGzLoeqzqHrt/xkw=">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</latexit>

d
X
x(k, l) = ai (k, l)si (k, l) + n(k, l)
i=1

• Assuming a single target and considering remaining point sources as

interferers, abusing notation we can write
d
X
x(k, l) = a1 (k, l)s1 (k, l) + ai (k, l)si (k, l) + n0 (k, l)
| {z }
i=2
target | {z }
interf erers+noise

= a(k, l)s(k, l) + n(k, l)

• Goal: Estimate s(k, l) given x(k, l): e.g. ŝ(k, l) = E[s(k, l)|x(k, l)]
 Summary of Previous Lecture
• Delay and sum beamformer
<latexit sha1_base64="sAs0bfPvvMBkqim1LI3db4FMCXI=">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</latexit>

a(k, l)
w(k, l) =
aH (k, l)a(k, l)

• MVDR beamformer
Rx 1 (k, l)a(k, l) Rn 1 (k, l)a(k, l)
w(k, l) = H 1 = H
a (k, l)Rx (k, l)a(k, l) a (k, l)Rn 1 (k, l)a(k, l)

• Multi-Channel Wiener
2
s (k, l) Rn 1 (k, l)a(k, l)
w(k, l) = 2
s (k, l) + (aH (k, l)Rn 1 (k, l)a(k, l)) 1aH (k, l)Rn 1 (k, l)a(k, l)
| {z }| {z }
Single-channel Wiener M V DR

6
 Optimal Linear Multi-Channel Wiener

<latexit sha1_base64="YegtaDd7GPm9QQuITcXeKy30gUw=">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</latexit>

Signal model: x(k, l) = s(k, l)a(k, l) + n(k, l)

Cost function: JM SE (w(k, l)) = E[||s(k, l) wH (k, l)x(k, l)||22 ]

dJM SE (w(k, l))

= E[sH (k, l)x(k, l)] + Rx (k, l)w(k, l)
dwH (k, l)
2
= s (k, l)a(k, l) + Rx (k, l)w(k, l) = 0

w(k, l) = Rx 1 (l) 2
S,k a(k, l)

7
 Optimal Linear Multi-Channel Wiener

Using again the Matrix inversion lemma, it can be shown that

<latexit sha1_base64="o6ik9cvoxBTKWHeLMwC5eMc+CE4=">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</latexit>

w(k, l) = Rx 1 (k, l) 2
s (k, l)a(k, l)

can be written as
2
s (k, l) Rn 1 (k, l)a(k, l)
w(k, l) = 2
s (k, l) + (aH (k, l)Rn 1 (k, l)a(k, l)) 1 aH (k, l)Rn 1 (k, l)a(k, l)
| {z }| {z }
Single-channel Wiener M V DR

8
 Optimal Linear Multi-Channel Wiener

matrix inversion lemma:

<latexit sha1_base64="hhyqroeKUrnMIC/FnqWwE3yLIpI=">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</latexit>

T 1 1 A 1 uvT A 1
(A + uv ) =A
1 + vT A 1 u

Matrix Rx (k, l) can be written as Rx (k, l) = Rn (k, l) + aaH 2

s (k, l)

1
Rx 1 (k, l)a(k, l) 2
s (k, l) = Rn (k, l) + aaH 2
s (k, l) a(k, l) s2 (k, l)
= Rn 1 (k, l)a(k, l) s2 (k, l)
2 H 1
1 s (k, l)a (k, l)Rn (k, l)a(k, l) 2
Rn (k, l)a(k, l) 1 s (k, l)
1 + s2 (k, l)aH (k, l)Rn (k, l)a(k, l)
✓ 2 H 1
◆
s (k, l)a(k, l) R n (k, l)a(k, l)
= Rn 1 (k, l)a(k, l) 1 2
s (k, l)
1 + s2 (k, l)aH (k, l)Rn 1 (k, l)a(k, l)

9
 Optimal Linear Multi-Channel Wiener

<latexit sha1_base64="+I1E4lGp+rsgFK0u1Sit6cgNj5Y=">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</latexit>

✓ 2 H 1
◆
1 s (k, l)a(k, l) R n (k, l)a(k, l) 2
= Rn (k, l)a(k, l) 1 2 H 1 s (k, l)
1 + s (k, l)a (k, l)Rn (k, l)a(k, l)
✓ 2
◆
1 s (k, l)
= Rn (k, l)a(k, l)
1 + s2 (k, l)aH (k, l)Rn 1 (k, l)a(k, l)
1
✓ H 1 2
◆
Rn (k, l)a(k, l) a (k, l)Rn (k, l)a(k, l) s (k, l)
=
aH (k, l)Rn 1 (k, l)a(k, l) 1 + s2 (k, l)aH (k, l)Rn 1 (k, l)a(k, l)
!
Rn 1 (k, l)a(k, l) 2
s (k, l)
=
aH (k, l)Rn 1 (k, l)a(k, l) aH (k, l)Rn 1 (k, l)a(k, l)
1
+ 2
s (k, l)

10
 Optimal Linear Multi-Channel Wiener
2
Rn 1 (k, l)a(k, l)
<latexit sha1_base64="FrcMOoW/P3TFlWlKtLy4PxrtC8A=">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</latexit>

s (k, l)
w(k, l) = 2
s (k, l) + (aH (k, l)Rn 1 (k, l)a(k, l)) 1aH (k, l)Rn 1 (k, l)a(k, l)
| {z }| {z }
Single-channel Wiener M V DR

The multi-channel Wiener filter can thus be seen as a concatenation

of two filters:

• An MVDR as spatial filter

• Single-Channel Wiener filter as post-processor where the noise
variance is set to the remaining noise PSD after beamforming:

wH (k, l)Rn (k, l)w(k, l) = aH (k, l)Rn 1 (k, l)a(k, l)

11
 <latexit sha1_base64="JvZWNMkMqno+z+0a+/BqmtcJj1A=">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</latexit>
Example: Multi-Channel Noise Reduction
x1 (k, l)
Passing car Babble noise
(120 degrees) (80 degrees) Target
90
60(0 degrees)
<latexit sha1_base64="2wU2EZvGwP5bNhSPlUSEDcejTqg=">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</latexit>

<latexit sha1_base64="sKwXBYJcRggo5MXaLqltbx8ob7E=">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</latexit>

z(k, l) ŝ(k, l) 120

1.5
Single-channel
MVDR- MMSE
beamformer Noise reduction
1
<latexit sha1_base64="j0qFu45Z6ZKnZ09MTujREntfGxY=">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</latexit>
150 30
xM (k, l)
0.5

180 0

• Clean input s(k, l)

<latexit sha1_base64="Trm/II/byVEk1gFIZMvGQ7oVBh0=">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</latexit>

• Noisy input at mic. 1 x1 (k, l) 210 330

• MVDR-beamformer output z(k, l)

240 300
• Output of total system ŝ(k, l) 270
12

May 30, 2023

 Sufficient Statistics
• For n Gaussian distributed,
<latexit sha1_base64="5/n+22JLqyaimFhy62VvLwPjzjU=">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</latexit>

H 1
H a (k, l)R n (k, l)x(k, l)
T (x(k, l)) = wMVDR (k, l)x(k, l) = H
a (k, l)Rn 1 (k, l)a(k, l)

is known to be a sufficient statistic for s.

• This means no information is lost on s by using T (x(k, l))
instead of x(k, l).
• This result holds in general for any prior distribution on s(k, l)
and any cost-function (e.g., quadratic (MSE), uniform (MAP),
Absolute error (Median)) and any function of s (e.g., |s|, |s|2 ,
etc.)
13
 Sufficient Statistics
• Let fS (s|x) denote the conditional pdf of random variable S. It
<latexit sha1_base64="A+Eh8VlyS3LPnsGDWUjz2L1AMbw=">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</latexit>

then holds that for a sufficient statistics fS (s|x) = fS (s|T (x)).

• If fx (x|T (x; s) is independent of s, T (x) is a sufficient statistic
for estimating s.

• Equivalent: I (s; T (x)) = I (s; x), i.e., we have equality in the

data processing inequality and no information is lost.
Finding a sufficient statistic: if the pdf fx (x; s) can be factorized as

fx (x; s) = g(T (x), s)h(x),

then T (x) is a sufficient statistic for s.

14
 LCMV - beamformer

Remember the MVDR: J(w(k, l)) = wH (k, l)Rx (k, l)w(k, l)

<latexit sha1_base64="BkPfWFIIM6Yh8v/kM3mg8JJflD0=">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</latexit>

min J(w(k, l))

w(k,l)

s.t.w(k, l)H a(k, l) = 1.

• The MVDR imposes one constraint.

• This can be generalised to having d constraints.

15
 LCMV - beamformer

Cost function: J(w(k, l)) = wH (k, l)Rx (k, l)w(k, l)

<latexit sha1_base64="8CDdAwEEMMP+oOWPqFFsoCh7cIc=">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</latexit>

min J(w(k, l))

w(k,l)

s.t.wH (k, l)⇤(k, l) = f H (k, l).

with ⇤ 2 CM ⇥d

When d < M , there is a closed form solution:

<latexit sha1_base64="yRrka5Dj7k1SNyegyCExlETpSaM=">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</latexit>

1 H 1 1
w(k, l) = Rx (k, l)⇤(k, l) ⇤ (k, l)Rx (k, l)⇤(k, l) f (k, l).

16
 LCMV - beamformer

1
<latexit sha1_base64="jw28MbXiGToO1d+z9BVbP4VUWvs=">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</latexit>

1 H 1
wk = Rx (k, l)⇤(k, l) ⇤ (k, l)Rx (k, l)⇤(k, l) f (k, l).
How to use the multiple constraints?
• To steer zeros in the direction of certain noise sources.

• To maintain the signal from certain directions.

• To maintain the spatial cues of for hearing aids.
Notice that the more constraints are used, less degrees of freedom
are left to control the noise reduction.

17
 Overview of Discussed filters
• Delay and sum beamformer
<latexit sha1_base64="sAs0bfPvvMBkqim1LI3db4FMCXI=">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</latexit>

a(k, l)
w(k, l) =
aH (k, l)a(k, l)

• MVDR beamformer
Rx 1 (k, l)a(k, l) Rn 1 (k, l)a(k, l)
w(k, l) = H 1 = H
a (k, l)Rx (k, l)a(k, l) a (k, l)Rn 1 (k, l)a(k, l)

• Multi-Channel Wiener
2
s (k, l) Rn 1 (k, l)a(k, l)
w(k, l) = 2
s (k, l) + (aH (k, l)Rn 1 (k, l)a(k, l)) 1aH (k, l)Rn 1 (k, l)a(k, l)
| {z }| {z }
Single-channel Wiener M V DR

18
 Overview of Discussed filters

• LCMV beamformer
<latexit sha1_base64="UlFgxPYt0Omrenq5ffLYj4R/LBw=">AAAFznicnVRNb9MwGE4ZhRG+NjhyqWgnbVIWJam6sQPSRA/ssEljWreJuStO4qRW46TY7tZiWdwQN/hB/BH+CifsNNVoyiSGVVXvx+Pn/XT8YYIZd5yflTtLd6v37i8/MB8+evzk6crqsxOWjWiAOkGWZPTMhwwlOEUdjnmCzoYUQeIn6NQftLX/9BJRhrP0mE+GqEtgnOIIB5ArU2/lF0gznIYo5SbwUYxTgTki+BOSJtCSud8+OKn5ijHKKEHUBJdsCAMkNh27GRCFOgcE8r4fiSu5PrCSjdcCKOVIygux6cqeyNWxlFNvroF9lWAICxNIUMTX5x0Xe9fo21BRHPf5Rg6fpRVNXTbomgCl4XV9vZW6Yzv5qS0KbiHUjeIc9laXfoAwC0ZEdStIIGMCUo6DRHGtmWDEkGrMAMZIQMKiLOXMUoLOQs5586HNm9CQRTjOaShK0VWQEQJVriCCBCeTEEVwlHApAItmsgkYUtNOY94XgKMxv8KhiiRcR09FEf1Jo6Dy3O2KaU98UXelnEccSZHDdLdLrvc3u9ozV1u7SkEDTtVqFYG1BPrF6tg7LTSWgELMkJ+NtUHpogH6kAvZkNfAbULUIDUG0FGCzjebhHSFo6z5n9oJVQqwSmkFyeCGyJ7t/lNk97/ixhSGRZ/11QZIoZ/AnijVkzN69jyl5itgdkvpjXKvNTn7SBf5L7zbR2gWEUojCzOiJ0pJTUu6vjVzraZWTn1davneMn1lash1cTA5QqEUNPalsLcsx3JkGfAmGaEC4VgKY79agLylCKUzFk8hLNv7G2gyw7Ss/FfCHHdKsVqu4mrueLoQ9d7d8uteFE7UimzZ7juvvrtVvPxl44Xx0lg3XGPb2DX2jEOjYwSVD5UvlW+V79XD6mVVVj9PoXcqxZ3nxtypfv0N1rcCDA==</latexit>

1
w(k, l) = Rx 1 (k, l)⇤(k, l) ⇤H (k, l)Rx 1 (k, l)⇤(k, l) f (k, l).

19
 Today

All beamformers discussed so far depend on the acoustic transfer

<latexit sha1_base64="BXPSlaSSybriOtAUZj19htVRjj8=">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</latexit>

function (ATF) a.

Today:
• How to estimate a

• How to estimate RS (rank-r extention)

• Beamformers in terms of RS

20
 Acoustic transfer function

The acoustic events in a room, under some assumptions, can be math-

<latexit sha1_base64="iCvKq06vvCS/OlQeXHn6WwPNANA=">AAAGCnicjVNNb9tGEGWUqk3Yr6Q95rKoZcAuaEJSYSdFUSCtD62BFEgFOwniFYzlckgtvB/C7lK2QPAf9FfkJ+RQoOi1h17bW/9NZ0kpseUWCEEQw5m3772dnc3mUjg/HP5zq3f7vf77H9y5G3/40ceffHrv/mfPnKkshxNupLEvMuZACg0nXngJL+YWmMokPM/OD0P9+QKsE0Yf++UcpoqVWhSCM4+ps/u9b6g2QuegfXw8A8K4qZwXnMACU44ITRixxqiEVIiyxBmFKOcqNQ8MLiGcaZIBUczPAD9ILeWSiBwY+occwcSbgAgemSVM58QLBXtCL5gVTHuy8+T4aBepiZ8xjx/AGOXCUuHJhalk3koAc5UNlB3IAgeBu1tbQGVeSeYRkQssevSRLQk3emFkFewSU7QrJ0eTzkcnhb0kTpSayZTEdOHmjEOd7tN1Z925mDeUxt+FfcNbl57ZEvxqaUIcW5KBGyREOOzaHPvqV+xp/ARxg9mZGhDstfHwxkZhjbrqAwlCv5Tg1sxnRmNGDVJypDtZjpaSDt+qksFlIN0JGTxJB3uFhY1NrRv1pnHI6GdXJHaD4xIR2MZlTE9j5DzVU/ItoRIKv4O+yZfEUSvKmd/FSkyncRyf3dsapsP2ITeD0SrYilbPUxy3X2luOPZQey5ximpmcWAkNPF2TCsH2PhzVkLNlCsMzl+CQRis5lq1HfvrKZi7QpQtjQUNF9wohQdc04IpIZc5FKySvqmpK9ZxTB3gfdGln9XUw6W/EDkq1aNhOuaqpbpKhODmdDStafCTZfXWqGmuIyZN3cLqSbNZerkuvbxROlyXDkNpQ5R7G6aoEw4RnXXDuZd+vQ+XDbUMjzwzlyGB//WA4pDUzaB5C3yoVENVwFBbSTjd+0qpaT3EbPtpzsJWaLJhi8vz/1Eep6N3Uh69q+52vI2XosB7RtqTdaELXaL9r39aTiBvaltmTZ0eJMNk2GwCvpcVrBDDBDHpoxuQH/Be6DXLGBFJOv4v0HKN2U/aFw1GOOmjzbm+GTzD1hyko5/HW48PVjN/J3oQfRHtRKPoYfQ4+jF6Gp1EvPeq92fvr97f/V/6r/u/9X/voL1bqzWfR9ee/h//AhrjDGE=</latexit>

ematically idealised as to be linear and time-invariant (LTI) so that the

sound as it would be measured at the receiver can be calculated di-
rectly by convolution of the RIR and the source signal.
Assume that the target signal, say s, is a point source. Let hm denote
the RIR from the source s to microphone m. In that case, the signal
xm (the noise-free source signal received at the mth microphone) is
given by
xm [n] = (hm ⇤ s) [n]

May 30, 2023 21

 Acoustic transfer function

Using the STFT, we can write equivalently

<latexit sha1_base64="SZ1yvkDGZdc6eZNLsDco1DEOgFM=">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</latexit>

xm (k, l) = am (k, l)s(k, l)

• The function am (k, l) is called the acoustic transfer function

(ATF) from the source to the mth microphone.

• Notice that it is thus the temporal frequency domain represen-

tation of the room impulse response.
• Given a fixed frequency band k, we can collect the M microphone
T
DFT coefficients in a vector x(k, l) = [x1 (k, l), ..., xM (k, l)]
such that x(k, l) = a(k, l)s(k, l).

May 30, 2023 22

 Relative Acoustic transfer function

• In many applications we are interested in the relative acoustic

<latexit sha1_base64="E5Y5yUVEJe45mW2GQTrAXEBzngo=">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</latexit>

transfer function, which is then normalize with respect to a ref-

erence location, e.g., with respect to one of the microphones,
T
a0 (k, l) = [1, a2 (k, l)/a1 (k, l)..., aM (k, l)/a1 (k, l)] .

May 30, 2023 23

 Cross Power Spectral Density Matrices

<latexit sha1_base64="is13gmWXBaLgLTtvi40DhYluUgc=">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</latexit>

Assuming that all sources (si [n] and n[n]) are realizations of random
processes, we can define the cross power spectral density matrix per
frequency band k and time frame l:

E[x(k, l)xH (k, l)] = E[s(k, l)sH (k, l)] + E[n(k, l)nH (k, l)]
| {z } | {z }
target source interferers/noise

often written as

Rx (k, l) = Rs (k, l) + Rn (k, l)

May 30, 2023 24

 Cross Power Spectral Density Matrices
<latexit sha1_base64="ToUb98C8Zd9Q3X+R7JVfn2wz3mM=">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</latexit>

Assume that the target signal s[n] is a point source, so that

s(k, l) = a(k, l)s1 (k, l),

where a 2 CM is the (relative) acoustic transfer functions from the

source to the microphones. With this, Rs can be expressed as

Rs = E[s(k, l)s(k, l)H ] = 2 H

s1 a(k, l)a (k, l),

where s21 = E[|s1 (k, l)|]2 , the variance of the clean signal as received
at the reference microphone 1.

Note that in this case rank(Rs ) = 1.

May 30, 2023 25

 Cross Power Spectral Density Matrices
<latexit sha1_base64="sU/kLr5D+OzsFv+jd1N6iMEXXVY=">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</latexit>

The beamformers derived so far assume that the (relative) acoustic

transfer function a(k, l) is known a-priori
• In practice, a(k, l) is unknown and needs to be estimated
• Estimation errors in a(k, l) generally lead to severe performance
degradation of the beamformer
• When there are multiple sources, the beamformers will be a func-
tion of a general correlation matrix Rs
In the following, we will focus on 1) estimating a, 2) estimating Rs
and 3) give expressions for beamformers in terms of a general Rs , not
necessarily of rank 1

May 30, 2023 26

 The Eigenvalue Problem

The eigenvalue problem for a square matrix A is

<latexit sha1_base64="ceEsSjr1b1pO6F5tw/Ut4637z2A=">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</latexit>

Ax = x , (A I) = 0.
Any that makes A I singular is called an eigenvalue, the corre-
sponding x is the eigenvector. It has an arbitrary norm, usually set
equal to 1.

We can collect the eigenvectors in a matrix:

2 3
1
6 7
A [x1 x2 . . .] = [x1 x2 . . .] 4 2 5 , AT = T⇤.
..
.

May 30, 2023 27

 The Eigenvalue Problem

Assume A is Hermitian.
<latexit sha1_base64="QOsoWePuUfopL7yiYdFSbwJkLVY=">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</latexit>

• Every eigenvalue is real:

kxk2 = xH x = xH Ax = xH AH x = xH x = kxk2

• The eigenvectors are mutually orthogonal (T 1

= TH ):
Let x and y be eigenvectors of A corresponding to distinct
eigenvalues and µ, respectively. Then
yH x = yH Ax = yH AH x = µyH x = µyH x
Hence ( µ)yH x = 0 and thus x ? y.

May 30, 2023 28

 The Eigenvalue Problem

If, in addition, A is positive semi-definite (xH Ax 0 for all x 2 Cn ),

<latexit sha1_base64="7wTyypDc3V6HXxm1NeF2ZNwAvlI=">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</latexit>

then
• Every eigenvalue is nonnegative:

kxk2 = xH x = xH Ax 0 ) 0

May 30, 2023 29

 The Eigenvalue Problem

If it exists, the eigenvalue decomposition of a square matrix A can be

<latexit sha1_base64="jfgNqi7EpF6weer+qAqN+nSGebk=">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</latexit>

written as
A = T⇤T 1 ,
with T invertible and ⇤ diagonal.

Notice that if A is Hermitian, that the eigenvalue are real and non-
negative, and the eigenvectors are orthogonal. If the eigenvectors have
norm one,T is unitary (T 1 = TH ) and thus

A = T⇤TH .

May 30, 2023 30

 Calculating the ATF - No Noise

Assume that Rs is perfectly known, i.e.,

<latexit sha1_base64="oaZLgAy634mqY11oSAAGTpQ5/ZM=">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</latexit>

Rs = E[s(k, l)s(k, l)H ] = 2 H

s1 (k, l)a(k, l)a (k, l).

Since Rs is Hermitian and positive semi-definite, there exists a unitary

matrix U = (u1 , . . . , uM ), ui 2 CM , such that

Rs = U⇤U 1
= U⇤UH ,

where ⇤ = diag( 1, . . . , M ), i 0 for all i.

In this case, ⇤ = diag s21 (k, l)ka(k, l)k2 , 0, . . . , 0 and the (scaled)
ATF is given by u1 = a(k, l)/ka(k, l)k

May 30, 2023 31

Calculating the ATF – Spatially White Noise
2
<latexit sha1_base64="co1NI1nT1QqlB5G8i6HrKgXMz5I=">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</latexit>

Now suppose Rx (k, l) is known and Rn (k, l) = n (k, i)I, i.e.,

2 H 2
Rx (k, l) = Rs (k, l) + Rn (k, l) = s1 (k, l)a(k, l)a (k, l) + n (k, i)I

As an identity matrix is diagonalizable by any unitary matrix, we have

2
Rx = Rs + nI
H 2
= U⇤U + nI

= U⇤UH + 2
n UU H

2 H
= U(⇤ + n I)U

which is the eigenvalue decomposition of Rx .

May 30, 2023 32

 Calculating the ATF
Conclusions:
<latexit sha1_base64="g4bjqH3AdbpqqFowJvw0QIcP+xk=">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</latexit>

• For spatially white noise, Rs (k, l) (which is not available) and

Rx (k, l) (which we can estimate from the observed data) share
the same eigenvectors

• Adding (spatially uncorrelated) noise to the desired speech data

only a↵ects the eigenvalues of Rs (k, l)
• Given Rx (k, l), assuming that the noise is spatially white, the
ATF can still be calculated by taking the principle eigenvector.
Remark: Notice that from now on, indices (k, l) will be neglected
because of notational convenience.
May 30, 2023 33
 Estimating Rs - Spatially White Noise
<latexit sha1_base64="i9iG90pa4L+LxNwUnFOlWGmnQF8=">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</latexit>

Let us assume that rank(Rs ) = r < M . We can partition Rx as

! !
⇤1 + n2 Ir O UH 1
Rx = U 1 U 2 2 H
,
O n IM r U2

where U1 2 CM ⇥r , U2 2 CM ⇥(M r)
and ⇤1 2 Cr⇥r .

Since Rx = U1 (⇤1 + n2 Ir )UH 1 + 2

n U 2 U H
2 , we conclude that the
eigenvectors u1 , . . . , ur span the speech (+ noise) subspace, whereas
ur+1 , . . . , uM span the noise only subspace.

Since U is unitary, we have UH

1 U2 = O (orthogonal subspaces).

May 30, 2023 34

 Geometric interpretation

(noise-only subspace) (speech + noise subspace)

U2 U1

May 30, 2023 35

 Estimating Rs - Spatially White Noise

Despite the fact that we do not know what the signal subspace is a
<latexit sha1_base64="+gToVdygdT7Yo2JeG4Oq67lnaek=">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</latexit>

priori (Rs is unknown), we can compute (estimate) Rs from the EVD

of Rx .

Least-squares estimate is obtained by approximating Rs by

R̂s = arg min kRx Rs k2F

rank(Rs )=r

The solution is a classical result and follows by truncating the M r

smallest eigenvalues. That is,
2 H
R̂s = U1 (⇤1 + I
n r )U 1

May 30, 2023 36

 Estimating Rs - Spatially White Noise

Since the last M r eigenvalues are given by n2 , we can even do

<latexit sha1_base64="pkgdaO9vL23QOMnI/JVOJZfbn4Q=">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</latexit>

better by subtracting n2 from the largest r eigenvalues (results in a

minimum variance estimator). That is,

R̂s = U1 ⇤1 UH
1 ,

which is identical to Rs .

<latexit sha1_base64="0fvcgiiKovyw+sm74wTBxtQCats=">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</latexit>

Note that in practice we have to estimate U and ⇤ (and thus U1

and ⇤1 ) from the noisy observations and for that reason the resulting
estimator is not identical to Rs although the above equation suggests
so.
May 30, 2023 37
 Estimating Rs – Pre-Whitening
<latexit sha1_base64="5vDFhr854BNEQ8QDHF9Qn7n3L3I=">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</latexit>

If the noise process n is not white (Rn 6= ↵IM for some ↵ > 0), we
can pre-whiten the data, assuming that Rn 0 (positive definite)

Since Rn is Hermitian and positive definite, we have

1 1 1 1 1 1
H H H H
Rn = U⇤U = U⇤ ⇤ U
2 2 = (U⇤ U )(U⇤ U ) = Rn Rn
2 2 2 2

1
where Rn is the (unique) Hermitian square root of Rn .
2

1
Consider the transformed process ñ = Rn n. The process ñ is spa-
2

tially white:
1 1
H H
Rñ = E(ññ ) = Rn E(nn )Rñ
2 2
= IM

May 30, 2023 38

 Estimating Rs – Pre-Whitening
1
<latexit sha1_base64="rlG2vXIJeQgqXeB2VSpW5AaRkEU=">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</latexit>

Next consider the transformed process x̃ = Rn x. Since this trans- 2

formation transforms the original noise process into a spatially uncor-

related one, we have
1 1 1 1
H H
Rx̃ = E(x̃x̃ ) = Rn E(xx )Rn
2 2
= R n R s R n + IM .
2 2

Hence, we can apply the same techniques as discussed previously to

the transformed process x̃ and de-whiten the result thus obtained.

May 30, 2023 39

 Estimating Rs – Pre-Whitening
Estimation of Rs :
<latexit sha1_base64="witxpQUylxdj48CrJhcfCr4fdKU=">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</latexit>

1 1
1. Compute Rn and pre-whiten the data: x̃ = Rn x
2 2
⇣ ⌘
2. Compute the EVD Rx̃ = Ũ ⇤ ˜ + IM ŨH , truncate the M r
smallest eigenvalues and reduce the remaining ones by one.
˜ 1 ŨH
3. Estimate R̂s̃ = Ũ1 ⇤ 1

4. De-whiten the result thus obtained so that

1 1
˜ H
R̂s = Rn Ũ1 ⇤1 Ũ Rn2
2
1

<latexit sha1_base64="kRQuTk1U07u7IO6mL0EKg9t9y/U=">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</latexit>

If rank (Rs (k, l)) = 1, the ATF for spatially non-white noise can thus
be obtained by selecting the principle eigenvector from R̂s or from
1/2
Rn Ũ1
May 30, 2023 40
 Generalised eigenvalue decomposition
Remarks:
<latexit sha1_base64="ZS8JFZGviV2dZeLVZxIOOAiXcBA=">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</latexit>

1
• The explicit use of Rn may result in a loss of accuracy in the
2

data
• Can be avoided by working directly with Rx and Rn

• In addition, when Rn and/or Rx are updated in a recursive way,

it is generally very complicated to update Rx̃ , while it is much
simpler to calculate updates of Rn 1 (using the matrix inversion
lemma)
Another (in theory equivalent) method is given by the generalised
eigenvalue decomposition

May 30, 2023 41