Anda di halaman 1dari 42

Adaptive Filtering Techniques for Forensic


Lino Garca Morales1, Jon Ander Beracoechea2, Soledad Torres-Guijarro3

and Yolanda Blanco-Archilla1
Universidad Europea de Madrid
Departamento de Electronica y Comunicaciones
E-mail: {lino.garcia,}
Universidad Politecnica de Madrid
Senales, Sistemas y Radiocomunicaciones
Universidad de Vigo
Departamento de Teora de Senal y Comunicaciones

Summary. The adaptive ltering techniques have plenty of applications in any ar-
eas where the modeled signals or systems are constantly changing. An adaptive lter
is a system whose structure is alterable or adjustable in such a way that its behavior
or performance improves through contact with its environment [16]. This chapter
focuses on adaptive ltering techniques for forensic audio applications. Multichan-
nel multirate specialized structures are presented as general cases. Five approaches
are studied: spectral equalization, adaptive linear prediction (ALP), adaptive noise
cancellation (ANC), beamforming and deconvolution or derreverberation. Objec-
tive and subjective measurements for the evaluation of intelligibility after speech
enhancement are revised.

1 Introduction

Forensic audio uses science and technology in order to assist a historical,

criminal and civil investigation in courts of law. There are sophisticated tech-
nologies and scientic methodologies of forensic audio that include several
specialties: voice and sound identication, audibility analysis, audio enhance-
ment, authenticity analysis and others. The employment of audio recordings of
telephone conversations and interviews, as well as covert surveillance record-
ings, is an integral part of law enforcement. Such recordings are frequently of
poor quality, but they are admissible if they are intelligible and meet the rules
of evidence. It can be equally important that the recording is listenable, and
the information that contains is easily discerned by the jury. Written tran-
scripts can be vital evidence, and these can only be made with condence if
2 Lino Garca et. al.

the recording is intelligible [5]. Forensic ltering can be used in the laboratory
to reject the noise and interference, as well as to restore, clarify or enhance the
audio information to assist law enforcement agencies criminal investigation,
civil investigation and the court process.
There is a number of possible degradations that can be found in a speech
recording and that can aect its quality. On one hand, the signal arriving
the microphone usually incorporates multiple sources: the desired signal plus
other unwanted signals generally termed as noise. On the other hand, there are
dierent sources of distortion that can reduce the clarity of the desired signal:
amplitude distortion caused by the electronics; frequency distortion caused by
either the electronics or the acoustic environment; and time-domain distortion
due to reection and reverberation in the acoustic environment.
Adaptive lters have traditionally found a eld of application in noise and
reverberation reduction, thanks to their ability to cope with changes in the
signals or the sound propagation conditions in the room where the recording
takes place. This chapter is an advanced tutorial about multichannel adaptive
ltering techniques suitables for forensic audio to provide relevant theoretical
foundation in this regard. The employment of more than one microphone is
useful for audio surveillance purposes. This is possible when the room where
the recording will be made can be prepared in advance. However, monochan-
nel adaptive ltering can be seen as a particular case of the more complex
and general multichannel adaptive ltering. The dierent adaptive ltering
techniques are presented in a common foundation useful in other forensic dis-
This chapter is organized as follows: in Sect. 1.1, we introduce a formal
denition of the forensic audio scenario from the multiple-input and multiple-
output (MIMO) perspective and the terminology that is used throughout the
chapter. In Sect. 1.2, signals and systems related to forensic audio are briey
summarized. Section 1.3 discusses quality measurements. Section 2 is dedi-
cated to the theoretical foundation of the adaptive lters. In Sects. 2.1 and 2.2
the lters structure and the adaptations algorithms are discussed. The dif-
ferent cost functions, stochastic estimations and optimization strategies over
transversal and lattice structures are shown. In Sect. 3 specialized structures
based on multirate techniques are presented. These schemes allow computa-
tionally ecient algorithms to be suitable for very large impulse responses
involved in forensic audio applications and real time implementations. Two
approaches are considered in Sects. 3.1 and 3.2: the subband and frequency-
domain partitioned adaptive ltering respectly. In Sect. 3.3 the partitioned
convolution is described and in Sect. 3.4 a delayless approach for real-time
applications are commented. Section 4 focuses on the adaptive ltering tech-
niques for forensic audio application: spectral equalization, linear prediction,
noise cancellation, beamforming and deconvolution.
Adaptive Filtering Techniques for Forensic Audio 3

1.1 General Issues and Definitions

Figure 1 shows the most general MIMO forensic audio scenario.

s1 (n) x1 (n) y1 (n)

V x2 (n) W y2 (n)
s2 (n)

xP (n) yO (n)
sI (n)


Fig. 1. Forensic audio scenario.

The box, on the left, represents a room where the evidence is being
recorded. V is a P LI matrix that contains the acoustic impulse responses
(AIR) between the I sources and P microphones 1

v11 v12 v1I
v21 v22 v2I

V= . .. . . .. ,
.. . . .
vP 1 vP 2 vP I

vpi = vpi1 vpi2 vpiL . (1)

Sources can be interesting or desired signals (to enhance) or noise and

interference signals (to attenuate). xp (n), p = 1 . . . P , is a corrupted or poor
quality signal that wants to be improved, (P = 1 corresponds to the sin-
gle channel case). r(n) is an additive noise or interference signal due to the
recording device. The forensic ltering goal is to obtain a W matrix so that
yo (n) si (n) corresponds to a restored or enhanced signal.
The signals in the Fig. 1 are related by

x(n) = Vs(n) + r(n), (2)

y(n) = Wx(n). (3)

s(n) is a LI 1 vector that collects the source signals,

Note that the discontinuous lines represent only the direct path and some rst re-
ections between the s1 (n) source and the microphone with output signal x1 (n).
Each vpi vector represents the AIR between i = 1 I and p = 1 P positions
and is constantly changing depending on the position of both: source or micro-
phone (i.e. mobile recording device), angle between them, radiation pattern, etc.
4 Lino Garca et. al.
s(n) = sT1 (n) sT2 (n) sTI (n) , (4)
si (n) = si (n) si (n 1) si (n L + 1) .
x(n) is a P 1 vector that corresponds to the convolutive system output
excited by s(n) and the adaptive lter input of order O LP . xp (n) is an
input corresponding to the channel p containing the last L samples of the
input signal x,
x(n) = xT1 (n) xT2 (n) xTP (n) , (5)
xp (n) = xp (n) xp (n 1) xp (n L + 1) .
W is an O LP adaptive matrix that contains an AIRs between the P
inputs and O outputs

w11 w12 w1P
w21 w22 h2P

W= . .. . . .. ,
.. . . .
wO1 wO2 hOP

wop = wop1 wop2 wopL . (6)
For a particular output o = 1 O, normally matrix W adapts in a rear-
ranged way like a column vector
w = w1 w2 w P . (7)
Finally, y(n) is an O 1 target vector,
y(n) = y1 (n) y2 (n) yO (n) .
The used notation is the following: a or is a scalar, a is a vector and A is a
matrix in a time-domain, a is a vector and A is a matrix in a frequency-domain.
Equations (2) and (3) are in matricial form and correspond to convolutions
in a time-domain. The index n is the discrete time instant related with the
time (in seconds) by means of a sample frequency Fs according to t = nTs ,
Ts = 1/Fs . Ts is the sample period. Superscript T denotes the transpose of
a vector or a matrix, denotes the conjugate of a vector or a matrix and
superscript H denotes Hermitian (the conjugated transpose) of a vector or a
matrix. Note that, if adaptive lters are a L 1 vectors, L samples have to
be accumulated per channel (i.e. delay line) to make the convolutions (2) and
From a signal processing point of view, the particular problem of noise re-
duction generally involves two major steps: modeling and ltering. The mod-
eling step generally involves determining some approximations of either the
noise spectrum or the input signal spectrum. Then, some ltering is applied to
emphasize the signal spectrum or attenuate/reject the noise spectrum. If the
parameters of the signal or noise model change over time, then the ltering
must be adaptive [7].
Adaptive Filtering Techniques for Forensic Audio 5

1.2 Signals and Systems

The involved signals in forensic audio science are very particular. In general, it
is possible to group them in three big classes: speech, noise, and the acoustic
impulse responses of the involved room.


Speech is produced by inhaling air into the lungs and exhaling it through a
vibrating glottis and the vocal tract. The random noise-like, air ow from the
lungs is spectrally shaped and amplied by the vibrations of the glottal cords
and the resonance of the vocal tract. The eect of the glottal cords and the
vocal tract is to introduce a measure of correlation and predictability on the
random variations of the air from the lungs [29]. The speech signal (voice)
is formed by silence, noisy and fricative segments and harmonic or occlusive
segments and is highly modulated.

Speech signal

0 0.1 0.2 0.3 0.4 0.5

400 0.8

200 0.2
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
Timefrequency analysis
2000 0

1000 40
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4

Fig. 2. Example of speech word. In the upper part a 0.5-seconds segment of speech
signal corresponding to the word sota [so ta] is depicted in the time-domain. The
signal was sampled at 8192 Hz. The superimposed solid light gray line corresponds
to the normalized energy of the segment and shows the low-frequency spectrum
that denes the rate at which we utter phonemes while the dotted dark gray line
corresponds to the normalized zero crossing. In the middle part, an autocorrelation
analysis of the speech signal is depicted. In the lower part a time-frequency analysis
of the same segment is depicted. Dark colors represent areas with high energy, light
colors display areas with low energy.

At the bottom of the Fig. 2 each one of these parts can be seen: rst a
fricative segment corresponding to s is a noisy segment, relatively low-pass;
although there is not a particular spectral region with preponderate energy.
6 Lino Garca et. al.

The next segment corresponding to o is clearly a harmonic segment. The

energy is concentrated in narrow frequency bands which correspond to the
formants and it is directly related with the autocorrelation (measure of pre-
dictibility) (in the middle of the Fig. 2). The next segment corresponding to
the silence has low energy contribution (probably due to recording device since
any anechoic chamber was used for recording). The next segment correspond-
ing to t is of very short duration and is weakly harmonic. The last occlusive
segment, corresponding the a phoneme, is clearly harmonic but with a dier-
ent contribution from o phoneme. The probability density function (pdf) of
the speech signal in time-domain is close to Laplacian and it can be assumed
to be sparse2.


The types of noises and interferences which can be present in the evidence
recording can be strong, subtle, and/or time varying and these may occur in
the acoustic environment where the microphones are located in. Some common
examples of acoustic noises and interferences include: air conditioning and fan
hum, reverberation and echoes, engines and other machinery, wind and rain,
radio and TV, live music, background speech in public places, other talkers,
vehicular trac and road noise, etc. Noise and unwanted sounds may lead to
listener fatigue and confusion for untrained listeners. Another class of noise
is related with a distortion that can be introduced by the recording devices
(bandwidth distortion). Both cases have a dierent approach. In the rst case
the noise and interference signal can be considered as any input signal si (n).
In the second case the noise r(n) is added equally to all the channels. Figure
3 shows an example of speech aected by an additive noise.


Understanding speech in a reverberant environment, such as a cathedral or

gymnasium can be very dicult. Reverberation is made up of sound reections
that have the eect of smearing, or blurring speech, making it less clear and
distinct and therefore more dicult to understand. This eect is reinforced as
the room reverberation time grows up. The energy of a reverberating signal
in a room depends on its size, and the materials inside it (dierent materials
having a dierent reection and absorption coecients) [12]. AIR or acoustic
transfer function is the relation between the receptor sound pressure and the
source sound pressure and can be modeled with direct sound, rst reections
and diuse eld. The AIR of a typical room can be very large (thousands
of taps) and in most cases is an unknown parameter. Figure 4 shows the
consequences of a reverberant environment.
A signal is sparse if only few of its samples are signicantly dierent from zero.
Sparseness is usually modeled by a Laplacian pdf.
Adaptive Filtering Techniques for Forensic Audio 7

Contaminated speech signal


0 0.1 0.2 0.3 0.4 0.5

400 0.8

200 0.2

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4

Timefrequency analysis

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4

Fig. 3. Example of a contaminated speech word. The same segment speech signal
depicted in the Fig. 2 is contaminated with a broad band noise signal with a -3 dB of
signal noise rate (SNR). It is dicult to recognize the occlusive segments, however,
in spite of the high level of the noise the speech signal is intelligible and perfectly
recognized (even until for SNR = -10 dB).

Reverberating speech signal


0 0.1 0.2 0.3 0.4 0.5

400 0.8

200 0.2

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
Timefrequency analysis
2000 0

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4

Fig. 4. Example of reverberating speech word. The same segment speech signal
depicted in the Fig. 2 is ltered simulating a recording of the evidence in the AIR
room. Note this the harmonic distortion.
8 Lino Garca et. al.

1.3 Target Quality

The goal in music restoration is to remove as much noise as possible without

introducing artifacts and to leave the remainder as musical as possible. This
is not constricted by these guidelines with forensic audio. The only goal, in
most cases, is to audibly discern what is being said by certain individuals so
that the message, previously obscured by the noise can now be heard and
understood. The audio enhancement is used to improve listenability and/or
intelligibility of a sound source by rejecting noise and interference signals (the
process otherwise known as restoration ). Noise reduction allows preparing the
evidence to minimize confusion without altering the nature of wanted signal.
Speech intelligibility is a measure of eectiveness of understanding speech,
as dened by the ISO standard [19]. The measurement is usually expressed as
a percentage of a message that is understood correctly. Speech intelligibility
does not imply speech quality. A synthesized voice message may be completely
understood by the listener, but maybe judged to be harsh, unnatural, and of
low quality. A message that lacks quality may still be intelligible.
Speech intelligibility can be assessed applying two dierent methods: sub-
jective assessment, based on the use of speakers and listeners; and objective
assessment, based on the measurement of physical parameters of the trans-
mission channel. Subjective tests are laborious, because they must involve a
number of speakers and listeners to assure representative results. Furthermore,
the results are dicult to reproduce, even if the test includes several refer-
ence conditions. Objective measurements are much faster and repeatable, but
their results may not be reliable, because they do not measure intelligibility
directly, but determine physical parameters to predict intelligibility accord-
ing to certain model. Such model might have restrictions that need to be
The subjective intelligibility measure may be based on phonemes, words
(meaningful or nonsense), and sentences; the results attained by using these
three types of material can be, in principle and under controlled conditions,
related to each other through the common intelligibility scale (CIS) [2]. The
use of sentences is advantageous in case of temporal distortion, such as the
introduced by reverberation, because they allow a closer simulation of this
kind of distortion on continuous speech. A very reproducible test, based on
sentence intelligibility, is the speech reception threshold (SRT) [24]. The SRT
would give an estimation of how much noise can the enhancement algorithm
cope with.
Behind objective intelligibility assessment underlies the assumption that
the intelligibility of a speech signal is based on the contribution of individual
frequency bands. In [9], French and Steinberg showed that the information
content of a speech signal is not equally distributed across frequencies, and
developed a model of twenty contiguous frequency bands that equally con-
tribute to a intelligibility index, the articulation index (AI). Based on this
model, several objective measurements have been developed for dierent ap-
Adaptive Filtering Techniques for Forensic Audio 9

plication elds. Among them, the only one which accounts correctly for band-
limiting noise, reverberation, echoes and non-linear distortion is the speech
transmission index (STI), standardized by IEC [18]. The STI is based on the
generation and analysis of an articial probe signal that replaces the speech
signal, on which it is easier to measure the eects of noise and distortion.
Under some circumstances, it is possible to mathematically derive the STI
from the impulse response of the transmission system, thus avoiding the use
of the probe signal. The applicability of STI to evaluate the performance of
speech enhancement algorithms will be discussed in Sect. 5.

2 Adaptive Filters. Why?

The major assumption in developing linear time-invariant (LTI) systems is
that the unwanted noise can be modeled by an additive Gaussian process.
However, in some physical and natural systems, noise can not be modeled
simply as an additive Gaussian process, and the signal processing solution
may also not be readily expressed in terms of mean squared errors (MSE)3
[7]. Adaptive ltering techniques are used largely in audio applications where
the ambient noise environment has a complicated spectrum, the statistics
are rapidly varying and the lter coecients must automatically change in
order to maintain a good intelligibility of the speech signal. Thus, ltering
techniques must be powerful, precise and adaptive.
The adaptive lter adjusts itself to remove the modeled signal represent-
ing the unwanted signal (noise and interference) while preserving the target
signal (speech) in such a way that the desired information can be recovered.
There are two major approaches depending on the number of available refer-
ence signals to help the estimation of the noise spectrum. Most non-referenced
noise reduction systems have only one single input signal. The task of esti-
mating the noise and/or signal spectra must then make use of the information
available only from the single input signal and the noise reduction lter will
also have only the input signal for ltering. Referenced adaptive noise reduc-
tion/cancellation systems work well only in constrained environments where
a good reference input is available, and the crosstalk problem is negligible or
properly addressed.
MSE is the best estimator for random (or stochastic) signals with Gaussian distri-
bution (normal process). The Gaussian process is perhaps the most widely applied
of all stochastic models: most error processes, in an estimation situation, can be
approximated by a Gaussian process; many non-Gaussian random processes can
be approximated with a weighted combination of a number of Gaussian densi-
ties of appropriated means and variances; optimal estimation methods based on
Gaussian models often result in linear and mathematically tractable solutions
and the sum of many independent random process has a Gaussian distribution
(central limit theorem) [29].
10 Lino Garca et. al.

The number of channels is related to the number of microphones recording

the evidence. More microphones are better than one. In a multichannel sys-
tem (P > 1) it is possible to remove noise and interference signals by applying
sophisticated adaptive ltering techniques that use spatial or redundant in-
formation. However there are a number of noise and distortion sources that
can not be minimize by increasing the number of microphones. Examples of
this are the surveillance, recording, and playback equipment (i.e. body wire
on a police informant, a telephone tap, a wireless transmitter, an emergency
service phone recorder, a memo recorder, shotgun microphone or parabolic
dish, video tape recorder, or the like).
There are several classes of adaptive ltering [17] that can be useful for
audio forensic, as will be shown in Sect. 4. The dierences amoung them are
based on the external connections to the lter.

x(n) Adaptive


x(n) Adaptive y(n) e(n)


(c) d(n)

x(n) Adaptive y(n) e(n)


Fig. 5. Classes of adaptive ltering.

In the estimator application [see Fig. 5(a)], the internal parameters of the
adaptive lter are used as estimate.
In the predictor application [see Fig. 5(b)], the lter is used to lter an
input signal, x(n), in order to minimize the size of the output signal, e(n) =
x(n) y(n), within the constrains of the lter structure. A predictor structure
is a linear weighting of some nite number of past input samples used to
estimate or predict the current input sample.
Adaptive Filtering Techniques for Forensic Audio 11

In the joint-process estimator application [see Fig. 5(c)] there are two
inputs, x(n) and d(n). The objective is usually to minimize the size of the
output signal, e(n) = d(n) y(n), in which case the objective of the adaptive
lter itself is to generate an estimate of d(n), based on a ltered version of
x(n), y(n) [17].

2.1 Filter Structures

Adaptive lters, as any type of lter, can be implemented using dierent

structures. There are three types of linear lters with nite memory: the
transversal lter, lattice predictor and systolic array [16].


The transversal lter, tapped-delay line lter or nite-duration impulse re-

sponse lter (FIR) is the most suitable and the most commonly employed
structure for an adaptive lter. The utility of this structure derives from its
simplicity and generality. Its transfer function can be changed easily con-
trolling the L coecients wl , l = 1 L. There is a simple linear relation-
ship between

L the parameters of the lter and the transfer lter function
W (z) = l=1 wl z l . In a FIR lter, each output value y(n) is determined
by a nite weighted combination of L previous values of the input signal

y(n) = wl x(n l + 1) = wH x(n) = w, x(n) , (8)

 T  T
with w = w1 w2 wL and x(n) = x(n) x(n 1) x(n L + 1) .
Equation (8) is called nite convolution sum 4 .

x(n) z 1 z 1 z 1

w1 w2 wL


Fig. 6. Transversal lter.

Although the audio signals are real, the Hermitian operator H is used (instead
of Transpose operator, T ) as the most general case because these structures can
be used within multirate specialized structures [Sect. 3] that generate complex
12 Lino Garca et. al.

The transversal structure in Fig. 6 is a rather direct realization of the

transfer function of W (z) in terms of delays and multipliers.
The transversal predictor lter output in Fig. 5(b) is the dierence between
the current input sample x(n) and the predicted value, given by

y(n) = wl x(n l) = w, x(n l) . (9)

The transversal lter output used to build a joint-process estimator as

illustrated in Fig. 5(c) is given by

y(n) = wl x(n l + 1) = w, x(n) . (10)

In both applications the error depends on w, e = f {w}.

Multichannel Adaptive Transversal Filtering Extension
Figure 7 shows a multichannel adaptive ltering scheme using a transversal


x1 (n) y1 (n) y(n) e(n)


xP (n) yP (n)

Fig. 7. Multichannel adaptive ltering.

The multichannel transversal lter output used to build a joint-process

estimator as illustrated in Fig. 5(c) is given by

y(n) = wpl xp (n l + 1) = wp , xp (n) = w, x(n) , (11)
p=1 l=1 p=1

where x(n) is dened in (5) and w in (7).

Adaptive Filtering Techniques for Forensic Audio 13


The lattice lter is an alternative to the transversal lter structure for the
realization of a predictor [10]. Figure 8 shows a Lattice-ladder joint-process
estimation consisting of L 1 stages; the number L 1 is refered to as the
predictor order. The coecient of the lattice structure, kl , l = 1 L 1, are
commonly called the reection or PARCOR coecients. In this framework,
instead of applying the input signal to a tapped-delay line, a prediction linear
lattice structure is used between both of them. The L observations in the x(n)
vector are replaced by the set of backward prediction errors b(n).

Stage 1 Stage L 1
x(n) f1 (n) f2 (n) fL1 (n) fL (n)
k1 kL1

b1 (n) k1 b2 (n)
bL1 (n) kL1 bL (n)
1 1
z z
w1 w2

Fig. 8. Multistage lattice lter.

The prediction problem is to nd the projection of x(n) on the subspace

spanned by x(n 1), x(n L). The lattice lter is a consequence of nding
a new set of vectors which also span this subspace, at the same time as they
have the valuable property of being mutually orthogonal. Considering (9) on
both prediction errors: forward and backward predictors, and that due to
the symmetry of the autocorrelation function, bl = fLl+1 , 1 l L, the
recursive equations of the prediction errors are given by

fl (n) = fl1 (n) + kl bl1 (n 1), f1 (n) = x(n), (12)

bl (n) = bl1 (n 1) + kl fl1 (n), b1 (n) = x(n). (13)

Applying the projection theorem and knowing that the optimal backward
2 2
and forward prediction errors have the same norm f l (n) = bl (n) , 1
l L, the reection coecients can be obtained as

bl1 (n 1), f l1 (n)

kl = . (14)
f l1 (n) bl1 (n 1)

A lattice predictor has a modular structure, which consists of a number of

individual stages, each of them has the appearance of a lattice [16].
14 Lino Garca et. al.

The joint-process estimation, of the lattice-ladder structure, is specially

useful for the adaptive ltering because its predictor, diagonalizes completely
the autocorrelation matrix. The transfer function of a lattice lter structure
is more complex than a transversal lter because the reexion coecients are

b(n) = Qx(n), (15)

y(n) = w b(n), (16)
where b(n) = b1 (n) b2 (n) bL (n) is a backward predictor vector, Q is a
lower triangular matrix that depends on the reection coecients as follows

1 0 0 0 0
k1 1 0 0 0

k1 k2 1 0 0

Q= . .. .. .. .. .. (17)
.. . . . . .

kL2 k1 kL2 k2 kL2 1 0
kL1 k1 kL1 k2 kL1 kL2 kL1 1
and x(n) = x(n) x(n 1) x(n L + 1) . There are L coecients of the
transversal part and L1 reection coecients. Q is a LL matrix which acts
like a preconditioner over the x(n) vector that generates a decorrelated signal
b(n). The error for both: predictor and joint-process estimator applications
illustrated in Fig. 5(b,c) respectly, depends on e = f {w, k}.

Multichannel Adaptive Lattice Filtering Extension

Figure 9 shows a multichannel adaptive ltering scheme using a lattice struc-

ture. The multichannel version of lattice-ladder structure [21] must consider
the interchannel relationship of the reection coecients by each stage l

f l (n) = f l1 (n) + Kl bl1 (n 1) , f 1 (n) = x(n), (18)

bl (n) = bl1 (n 1) + Kl f l1 (n) , b1 (n) = x(n), (19)
 T  T
where f l (n) = f1l (n) f2l (n) fP l (n) , bl (n) = b1l (n) b2l (n) bP l (n) ,
k11l k12l k1P l
 T k21l k22l k2P l

x(n) = x1 (n) x2 (n) xP (n) , and Kl = . .. . . . .
.. . . ..
kP 1l kP 2l kP P l
A compact equivalent representation of 15 and 16 is possible:

b(n) = Ab(n 1) + Kf 1 (n), (21)

y(n) = wAb(n 1) + wKf 1 (n), (22)
Adaptive Filtering Techniques for Forensic Audio 15

x1 (n) f11 (n) f12 (n) f1(L1) (n) f1L (n)

b11 (n) b12 (n) b1(L1) (n) b1L (n)

z 1 z 1 d(n)
w11 w12 w1(L1) w1L
y1 (n)

y(n) e(n)

xP (n) fP 1 (n) fP 2 (n) fP (L1) (n) fP L (n)

bP 1 (n) bP 2 (n) bP (L1) (n) bP L (n)

z 1 z 1
wp1 wp2 wp(L1) wpL

yP (n)

Fig. 9. Multichannel adaptive ltering with lattice-ladder joint-process estimator.

where w = wT1 wT2 wTL is a LP 1 vector of the joint-process
estimator coecients, wl = w1l w2l wP l .
b(n) = bT1 (n) bT2 (n) bTL is a LP 1 backward predictor coecients
vector. A is a LP LP matrix obtained with a recursive development of (18)
and (19)

0P P 0P P 0P P 0P P 0P P
IP P 0P P 0P P 0P P 0P P

K1 K2 I 0 0P P 0P P
K1 K3 K2 K3 0P P 0P P 0P P

A= .. .. .. .. .. .. . (23)
. . . . . .

K KL3 K KL3 0P P 0P P 0P P
1 2
K1 KL2 K2 KL2 IP P 0P P 0P P
K1 KL1 K2 KL1 KL2 KL1 IP P 0P P

IP P is a matrix with only ones in the main diagonal and 0P P is a

P P zero matrix. K = IP P K1 K2 KL1 is a LP P reection
coecients matrix.
An equivalence with a multichannel transversal lter transfer function is
possible by rewritting Q matrix (17) and x(n) used in (16) as
16 Lino Garca et. al.

IP P 0P P 0P P 0P P
K1 IP P 0P P 0P P

K2 K
K 0 0P P
1 2 P P
.. ,
Q = ... ..
. . (24)

KL3 K KL3 0P P 0P P
KL2 K1 KL2 IP P 0P P
x(n) = x(n)T x(n 1)T x(n L 1)T , (25)
with x(n) = x1 (n) x2 (n) xP (n) . Note that (15), Q dened in (24),
is the multichannel version of Gram-Schmidt orthogonalization algorithm. The
determinant of Q matrix is unitary, therefore it is non singular and it has

2.2 Adaptation Algorithms

Once a lter structure has been selected, an adaptation algorithm must also be
chosen. From control engineering point of view, the forensic ltering is a sys-
tem identication problem that can be solved by choosing an optimum criteria
or cost function J(w) in a block or recursive approach. Several alternatives
are available, and they generally exchange increased complexity for improved
performance (speed of adaptation and accuracy of the transfer function after
adaption or misalignment5 ).

Cost Functions

Cost functions are related to the statistics of the involved signals and depends
on some error signal

J(w) = f {e(n)}. (26)

The error signal e(n) depends on the specic structure and the adaptive
ltering strategy but it is usually some kind of similarity measure between
the target signal si (n) and the estimated signal yo (n) si (n) (for I = O).
The most habitual cost functions are listed in Table 1.

Stochastic Estimation

Non-recursive or block methods process in lots for a transversal lter structure.

The input signal is divided into time blocks, and each block is processed
independently or with some overlap. This algorithms have a nite memory.
A misalignment for transversal lter is dened by  = v w2 / v2 . A con-
version between lattice and tranversal lter is possible and useful to setup an
equivalent misalignment measure for lattice structures.
Adaptive Filtering Techniques for Forensic Audio 17

Table 1. Cost functions for adaptive ltering

J(w) Comments
e (n) Mean squared error (MSE). Statistic mean operator

N n=0
e2 (n) MSE estimator. MSE is normally unknown
e2 (n) Instantaneous squared error

Absolute error. Instantaneous module error
n nm 2
e (m) Least squares (Weighted sum of the squared error)
E{f l (n)2 + bl (n)2 } Mean squared predictor errors (for a lattice structure)

The use of memory (vectors or matrice blocks) improves the benets of

the adaptive algorithms because they emphasize the variations in the cross-
correlation between the channels. However, this requires a careful structuring
of the data, and they also increase the computational exigencies: memory and
processing. For a p channel, the input signal vector dened in (5) happens to
be a matrix of the form
Xp (n) = xTp (n N + 1) xTp (n (N 1) + 1) xTp (n) , (27)

xp (n N + 1) xp (n (N 1) + 1) xp (n)
xp (n N ) xp (n (N 1)) xp (n 1)

Xp (n) = .. .. .. .. ,
. . . .
xp (n N L + 2) xp (n (N 1) L + 2) xp (n L + 1)
d(n) = d(n N + 1) d(n (N 1) + 1) d(n) , (28)

where N represents the memory size. The input signal matrix to the mul-
tichannel adaptive ltering has the form
X(n) = XT1 (n) XT2 (n) XTP (n) . (29)

In the most general case (with order memory N ), the input signal X(n) is
a matrix of size LP N . For N = 1 (memoryless) and P = 1 (single channel)
(29) is reduced to (5).
There are adaptive algorithms that use memory N > 1 to modify the
coecients of the lter, not only in the direction of the input signal x(n),
but within the hyperplane spanned by the input vector x(n)
and its N 1
immediate predecessors x(n) x(n 1) x(n N + 1) per channel. The
block adaptation algorithm updates its coecients once every N samples as

w(m + 1) = w(m) + w(m), (30)

w(m) = arg minJ(w).

The matrix dened by (27) stores K = L + N 1 samples per channel.

The time index m makes reference to a single update of the weights from time
n to the n + N based on the K accumulated samples.
18 Lino Garca et. al.

The stochastic recursive methods, unlike the dierent optimization deter-

ministic iterative algorithms, allow the system to approach the solution with
the partial information of the signals every time using the general rule

w(n + 1) = w(n) + w(n), (31)

w(n) = arg minJ(w).

The new estimator w(n+ 1) is updated from the previous estimation w(n)
plus adapting-step or gradient obtained from the cost function minimization
J(w). These algorithms have an innite memory. The trade-o between con-
vergence speed and the accuracy is intimately tied to the length of memory
of the algorithm.
The error of the joint-process estimator using a transversal lter with
memory can be rewritten like a vector as

e(n) = d(n) y(n) = d(n) XH (n)w. (32)

The unknown system solution, applying the MSE like the cost function,
leads to the normal or Wiener-Hopf equation. The energy of the error vector
(sum of the squared elements of the error vector) is given by the inner vector
product as6
J(w) = e = eH e = (d XH w)H (d XH w), (33)
= 2Xd + 2XXH w. (34)
The Wiener lter coecients are obtained by setting the gradient of the
square error function to zero, this yields
w = XXH Xd = R1 r. (35)

R is a correlation matrix and r is a cross-correlation vector dened by

X1 X1 X1 X2 X1 XP
X2 X1 X2 X2 X2 XP

R = XXH = . .. .. .. , (36)
.. . . .
r = Xd = X1 d X2 d XP d . (37)

Due to the nearness with which microphones are placed in scenario of

Fig. 1; for each i = 1 I input source, P (P 1)/2 relations are ob-
tained: xH H
p wq = xq w p for p, q = 1 P , with p = q. For vector u =

p=2 w p w 1 w 1 , it is possible to verify that Ru = 0P L1 , thus
The index time n is omitted by simplicity. All the vectors and matrices are refer-
enced at the same index time n.
Adaptive Filtering Techniques for Forensic Audio 19

R is not invertible and no unique problem solution exists. The adaptive al-
gorithm leads to one of many possible solutions which can be very dierent
from the target v. This is known as a non-unicity problem.
For a prediction application, the cross-correlation vector r must be slightly
modied assuming a particular form r = Xx(n 1), P = 1 and x(n 1) =
x(n 1) x(n 2) x(n N ) .
The optimal Wiener-Hopf solution wopt = R1 r requires the knowledge
of both magnitudes: the correlation matrix R of the input matrix X and the
cross-correlation vector r between the input vector and desired answer d. That
is the reason why it has little practical value. So that the linear system given
by (35) has solution, the correlation matrix R must be nonsingular.
It is possible to estimate both magnitudes according to the windowing
type of the input vector.
The sliding windowing method uses the sample data within a window of
nite length N . Correlation matrix and cross-correlation vector are estimated
averaging in time,

R(n) = X(n)XH (n)/N , (38)

r(n) = X(n)d (n)/N.

The method that estimates the autocorrelation matrix like in (38) with
samples organized as in (27) is known as the covariance method. The matrix
that results is positive semidenite but it is not Toeplitz.
The exponential windowed method uses a recursive estimation according
to certain forgetfulness factor in the rank 0 < < 1,

R(n) = R(n 1) + X(n)XH (n), (39)

r(n) = r(n 1) + X(n)d (n).

Applying the matrix inversion (Sherman-Morrison-Woodburys) lemma to

A1 = B1 B1 C I + DB1 C DB1 , (40)

where A = B + CD and B are assumed to be non-singular, the following

is yielded

R1 (n) = 1 R1 (n 1) (41)
X(n)XH (n)R1 (n 1)
2 R1 (n 1) .
I + 1 XH (n)R1 (n)X(n)
When the excitation signal to the adaptive system is not stationary and
the unknown system is time-varying, the exponential and sliding windowed
methods allow the lter to forget or to eliminate errors happened farther in
time. The price of this forgetfulness is a deterioration in the delity of the
lter estimation [12].
20 Lino Garca et. al.

A recursive estimator has the form dened in (31). In each iteration, the
update of the estimator is made in a w(n) direction. For all the optimiza-
tion deterministic iterative schemes, a stochastic algorithm approach exists.
It is enough to replace the terms related to the cost function and calculate
the approximate values by each new set of input/output samples. In general,
most of adaptive algorithms turn an optimization stochastic problem into a
deterministic one7 and the obtained solution is an approximation to the one
of the original problem. The stochastic approximation method to the system
identication problem yields to squared minimum estimation as

J(w) = eH e/N = (d XH w)H (d XH w)/N , (42)

w(n) = arg minJ(w).

The gradient g = J(w), dened in (34), can be estimated by means of

g = 2(r + Rw), or by the equivalent one g = Xe , considering R and r
according to (38) or (39). It is possible to dene recursive updating strategies,
per each l stage, for lattice structures as

Kl (n + 1) = Kl (n) + Kl (n), (43)

Kl (n) = arg minJ(Kl ).

Optimization strategies

Several strategies to solve w = arg minJ(w) are proposed [14] (usually of

the least square type). It is possible to use a quadratic (second order) approx-
imation of the error-performance surface around the current point denoted
w(n). Recalling the second-order Taylor series expansion of the cost function
J(w) around w(n), with w = w w(n), you have
J(w + w)
= J(w) + wH J(w) + wH 2 J(w)w (44)
Optimization deterministic iterative schemes require the knowledge of the
cost function, the gradient (rst derivatives) dened in (45) or the Hessian
matrix (second order partial derivatives) dened in (46,53) while stochastic
recursive methods replace these functions by impartial estimations.
J(w) J(w) J(w)
J(w) = w1 w2 wL
, (45)

2 J(w) 2 J(w) 2 J(w)
w1 w1 w1 w2 w1 wL
2 J(w) 2 J(w) 2 J(w)

w2 w1 w2 w2 w2 wL
2 J(w) = .. .. .. . (46)
. . .
2 J(w) 2 J(w) 2 J(w)
wL w1 wL w2 wL wL

Sampled data of the random variable are used.
Adaptive Filtering Techniques for Forensic Audio 21

The vector g(n) = J(w) is the gradient evaluated at w(n), and the
matrix H(n) = 2 J(w) is the Hessian of the cost function evaluated at
Several rst order adaptation strategies are: to choose a starting initial
point w(0), to increment election w(n) = (n)g(n); two decisions are due
to take: movement direction g(n) in which the cost function decreases fastest
and the step-size in that direction (n). The iteration stops when a certain
level of error is reached w(n) < ,
w(n + 1) = w(n) + (n)g(n). (47)
Both parameters (n), g(n) are determined by a cost function. The second
order methods generate values close to the solution in a minimum number of
steps but, unlike the rst order methods, the second order derivatives are
very expensive computationally. The adaptive lters and its performance are
characterized by a selection criteria of (n) and g(n) parameters.

Table 2. Optimization methods.

Method Denition Comments

SD (n) = gg
H Rg Steepest-Descent
CG (See below) Conjugate Gradient
NR (n) = Q Newton-Raphson

The optimization methods are useful to nd the minimum or maximum of

a quadratic function, like in (33). Table 2 summarized the optimization meth-
ods. SD is an iterative optimization procedure of easy implementation and
computationaly very cheap. It is recommended with cost functions that have
only one minimum and whose gradients are isotropic in magnitude respect
to any direction far from this minimum. NR method increases SD perfor-
mance using a weighting matrix carefully selected. The simplest form of NR
uses Q = R1 . Quasy-Newton methods (QN) are a special case of NR with
Q simplied to a constant matrix. The solution to (33) is also the solution
to the normal equation (35). The conjugate gradient (CG) [6] was designed
originally for the minimization of convex quadratic functions but, with some
variations, it has been extended to the general case. The rst CG iteration is
the same that the SD algorithm and the new successive directions are selected
in such a way that they form a set of vectors mutually conjugated to the Hes-
sian matrix (corresponding to the autocorrelation matrix, R), qH i Rqj = 0,
i = j. In general, CG methods have the form

gl l = 1,
ql = (48)
gl + l ql1 l > 1,
gl , ql 
l = , (49)
ql , gl pl 
22 Lino Garca et. al.

gl 2
l = 2, (50)
wl+1 (n) = wl (n) + l (n)ql . (51)

CG spans the search directions from the gradient in course, g, and a

combination of previous R-conjugated search directions. guarantees the
R-conjugation. Several methods can be used to obtain . This method (50)
is known as Fleetcher-Reeves. The gradients can be obtained as g = J(w)
and p = J(w g).

Table 3. Memoryless Least-Squares (LS) methods.

Method Denition Comments

LMS (n) = Least Means Squares

NLMS (n) = x(n) 2 + Normalized LMS

FNLMS (n) = p(n) Filtered NLMS
p(n) = p(n 1) + L(1 ) x(n)2
PNLMS (n) = xH (n)Qx(n)+ Proportionate NLMS

The memoryless LS methods in Table 3 use the instantaneous squared

error cost function J(w) = e2 (n). The descent direction for all is a gradient
g(n) = x(n)e (n). The LMS algorithm is a stochastic version of the SD op-
timization method. NLMS frees the convergence speed of the algorithm with
the power signal. FNLMS lters the signal power estimation; 0 < < 1 is a
weighting factor. PNLMS adaptively controls the size of each weight. Q is a
diagonal matrix that weights the individual coecients of the lters, is a re-
laxation constant and guarantees that the denominator never becomes zero.
These algorithms are very cheap computationally but their convergence speed
depends strongly on the spectral condition number of the autocorrelation ma-
trix R (that relate the extreme eigenvalues) and can get to be unacceptable
as the correlation between the P channels increases.
The projection algorithms in Table 4 modify the lters coecients in the
input vector direction and on the subspace spanned by the N 1 predecessors.
RLS is a recursive solution to the normal equation that uses MSE like cost
function8 . There is an alternative fast version FRLS. LMS-SW is a variant
of SD that considers a data window. The step can be obtained by a linear
search. APA is a generalization of RLS and NLMS. APA is obtained by pro-
jecting the adaptive coecients vector w in the ane subspace. The ane
subspace is obtained by means of a translation from the orthogonal origin to
the subspace where the vector w is projected. PRA is a strategy to reduce
the computational complexity of APA by updating the coecients every N
Calculated according to (41).
Adaptive Filtering Techniques for Forensic Audio 23

Table 4. Least-Squares with memory methods.

Method Denition Comments

RLS (n) = R1 (n) Recursive Least-Squares
g(n) = x(n)e (n)
LMS-SW (n) = gH (n)X(n)X H (n)g(n)+ Sliding-Window LMS
g(n) = X(n)e (n)

APA (n) = X(n)XH (n)+I
Ane Projection Algorithm
g(n) = X(n)e (n)
PRA w(n + 1) = w(n N + 1) + (n)g(n) Partial Rank Algorithm

(n) = X(n)XH (n)+I
g(n) = X(n)e (n)
DLMS (n) = x(n),z(n) Decorrelated LMS
g(n) = z(n)e (n)
z(n) = x(n) x(n),x(n1)
x(n 1)
TDLMS (n) = x(n)2 , QQ = I Transform-Domain DLMS
g(n) = x(n)e (n)

samples. DLMS replaces the system input by an orthogonal component to

the last input (order 2). This changes the updating vector direction of the
correlated input signals so that these ones correspond to uncorrelated input
signals. TDLMS decorrelates into transform domain by means of a Q matrix.
The adaptation of the transversal section of the joint-process estimator in
the lattice-ladder structure depends on the gradient g(n) and, indirectly, on
the reection coecients, through the backward predictor. Note that deriving
(16) respect to w, the gradient vector g(n) = b(n) is obtained. However, the
reection coecient adaptation depends on the gradient of y(n) with respect
to them
J(K) = J(K)K1
K2 J(K)
, (52)
2 2 2

J(K) J(K) J(K)
K 1 K1 K1 K2 1 KL
J(K) J(K) 2 J(K)
2 2

K2 K1 K2 K2 K2 KL
2 J(K) = .. .. .. . (53)
2 . . .
J(K) 2 J(K) 2 J(K)

In a more general case, concerning to a multichannel case, the gradient

matrix can be obtained as G = J(K). Two recursive updatings are necesary

wl (n + 1) = wl (n) + l (n)gl (n) (54)

Kl (n + 1) = Kl (n) + l (n)Gl (n) (55)

Table 5 resumes the least-squares for lattice.

24 Lino Garca et. al.

Table 5. Least-Squares for lattice.

Method Denition Comments

GAL l (n) = b (n) 2 Gradient Adaptive Lattice

gl (n) = e (n)bl (n)

l (n) = Bl1 (n)
Gl (n) = bl1 (n 1)f H H
l (n) + bl (n)f l1 (n 1)
CGAL (See below) CG Adaptive Lattice

GAL is a NLMS extension for a lattice structure that uses two cost func-
tions: instantaneous squared error for the tranversal part and prediction MSE
2 2
for the lattice-ladder part, Bl (n) = Bl (n1)+(1)(|f l (n))| +|bl (n 1))| ,
where and are relaxation factors. For CGAL, the same algorithm described
in (48-51) is used but it is necessary to rearrange the gradient matrices of
the lattice system in a column vector. It is possible to arrange the gradi-
ents of all lattice structures in matrices. U(n) = gT1 (n) gT2 (n) gTP (n)
is the P L gradient matrix with respect to the transversal coecients,
 T  T
gp = gp1 gp2 gpL , p = 1 P . V(n) = G1 (n) G2 (n) GP (n) is
a P (L 1)P gradient matrix with respect to the reection coecients; and
rearranging these matrices in one single column vector, uT vT is obtained
u = g11 g1L g21 g2L gP 1 gP L ,
v = G111 G1P 1 GP 11 GP P 1 G112 GP P (L1) .

gl l = 1,
ql = (56)
gl + l ql1 l > 1
uTl vTl l = 1,
gl =  T (57)
gl1 + (1 ) uTl vTl l>1
l = 2, (58)
wl+1 = wl + ul , (59)
Kl+1 = Kl + l Vl . (60)

The time index n has been removed by simplicity. 0 < < 1 is a forget-
fulness factor which weights the innovation importance specied in a low-pass
ltering in (57). The gradient selection is very important. A mean value that
uses more recent coecients is needed for gradient estimation and to generate
more than one conjugate directions vector (57).
Adaptive Filtering Techniques for Forensic Audio 25

3 Multirate Adaptive Filtering

The adaptive lters used for forensic audio are probably very large (due to the
AIRs). The multirate adaptive ltering works at a lower sampling rate that
allows reducing the complexity [26]. Depending on how the data and lters
are organized, these approaches may upgrade in performance and avoid end-
to-end delay. Multirate schemes adapt the lters in smaller sections at lower
computational cost. This is only necessary for real-time implementations. Two
approaches are considered. The subband adaptive ltering approach splits the
spectra of the signal in a number of subbands that can be adapted inde-
pendently and therefore the ltering can be carried out in a fullband. The
frequency-domain adaptive ltering partitions the signal in time-domain and
projects it into a transformed domain (i.e. frequency) using better proper-
ties for adaptive processing. In both cases the input signals are transformed
into a more desirable form before adaptive processing and the adaptive algo-
rithms operate in transformed domains, whose basis functions orthogonalize
the input signal, speeding up the convergence. The partitioned convolution is
necessary for fullband delayless convolution and can be seen as an ecient
frequency-domain convolution.

3.1 Subband Adaptive Filtering

The fundamental structure for subband adaptive ltering is obtained using

band-pass lters for basis functions and replacing the x gains for adaptive
lters. Several implementations are possible. A typical conguration uses an
analysis lter bank, a processing stage and a synthesis lter bank. Unfortu-
nately, this approach introduces an end-to-end delay due to the synthesis lter
bank. Figure 10 shows an alternative structure which adapts in subbands and
lters in full-band to remove this delay [25].
K is the decimation ratio, M is the number of bands and N is the prototype
lter length. k is the low rate time index. The sample rate in subbands is
reduced to Fs /K. The input signal per channel is represented by a vector
xp (n) = x(n) x(n 1) x(n L + 1) , p = 1 P . The adaptive lter
in full-band per channel wp = wp1 wp2 wpL is obtained by means of
the T operator as


wp = (hmK wpm )K gm , (61)


from the subband adaptive lters per each channel wpm , p = 1 P ,m =
1 M/2 [25]. The subband lters are very short, of length C = L+N 1

N  K
K + 1, which allows to use much more complex algorithms. Although the
input signal vector per channel xp (n) has size L 1, it acts as a delay line
which, for each iteration k, updates K samples. K is an operator that means
26 Lino Garca et. al.

z 1 z 1 z 1 d(n)


x1 (n) y1 (n) y(n) e(n)

e1 (k)
K w11
z 1
e2 (k)
K H w12
eM/2 (k)
z 1 K w1(M/2)

xP (n) yP (n)

wP 1
z 1
K H wP 2
z 1

z 1 wP (M/2)

Fig. 10. Subband adaptive ltering. This conguration is known as open-loop be-
cause the error is in the time-domain. An alternative closed-loop can be used where
the error is in the subband-domain. Gray boxes corresponds to ecient polyphase
implementations. See detail in [25].
Adaptive Filtering Techniques for Forensic Audio 27

downsampling for a K factor and K upsampling for a K factor. gm is a

synthesis lter in subband m obtained by modulating a prototype lter. H is
a polyphase matrix of a generalized discrete Fourier transform (GDFT) of an
oversampled (K < M ) analysis lter bank [8]. This is an ecient implemen-
tation of a uniform complex modulated analysis lter bank. This way, only a
prototype lter 9 p is necessary.
It is possible to select dierent adaptive algorithms or parameter sets for
each subband. For delayless implementation, the full-band convolution may
be made by a partitioned convolution.

3.2 Frequency-Domain Adaptive Filtering

The basic operation in frequency-domain adaptive ltering (FDAF) is to trans-
form the input signal in a more desirable form before the adaptation process
starts [26] in order to work with matrix multiplications instead of dealing with
slow convolutions. The frequency-domain transform employs one or more dis-
crete Fourier transforms (DFT), T operator in Fig. 11, and can be seen as a
pre-processing block that generates decorrelated output signals. In the more
general FDAF case, the output of the lter in the time-domain (11) can be
seen as the direct frequency-domain translation of the block LMS (BLMS)
algorithm. That eciency is obtained taking advantage of the equivalence
between the linear convolution and the circular convolution (multiplication
in the frequency-domain). It is possible to obtain the linear convolution be-
tween a nite length sequence (lter) and an innite length sequence (input
signal) with the overlapping of certain elements of the data sequence and the
retention of only a subgroup of the DFT.
The partitioned block frequency-domain adaptive ltering (PBFDAF) was
developed to deal eciently with such situations [23]. The PBFDAF is a more
ecient implementation of the LMS algorithm in the frequency-domain. It re-
duces the computational burden and bounds the user-delay. In general, the
PBFDAF is widely used due to its good trade-o between speed, computa-
tional complexity and overall latency. However, when working with long AIRs,
the convergence properties provided by the algorithm may not be enough. This
technique makes a sequential partition of the impulse response in the time-
domain prior to a frequency-domain implementation of the ltering opera-
tion. This time segmentation allows setting up individual coecient updating
strategies concerning dierent sections of the adaptive canceller, thus avoiding
the need to disable the adaptation in the complete lter. In the PBFDAF case,
the lter is partitioned transversally in an equivalent structure. Partitioning
wp in Q segments (K length) we obtain
Q K1
y(n) = xp (n qK m)wp(qK+m) . (62)
p=1 q=1 m=0
The prototype lter is a low-pass lter. The band-pass lters are obtained mod-
ulating a prototype lter.
28 Lino Garca et. al.


x1 (n) x1 (n) y11 (k)y1 (k) y(k) y(n) e(n)

T w11 T1
z K
y21 (k) T
yQ1 (k)

xP (n) xP (n) y1P (k)

T w1P
yP (k)
z K
y2P (k)
z K
yQP (k)

Fig. 11. Partitioned block frequency-domain adaptive ltering.

Where the total lter length L, for each channel, is a multiple of the length
of each segment L = QK, K L. Thus, using the appropriate data section-
ing procedure, the Q linear convolutions (per channel) of the lter can be
independently carried out in the frequency-domain with a total delay of K
samples instead of the QK samples needed by standard FDAF implementa-
tions. Figure 11 shows the block diagram of the algorithm using the overlap-
save method. In the frequency-domain with matricial notation, (62) can be
expressed as

Y = X W, (63)

where X = FX represents a matrix of dimensions M QP which contains

the Fourier transform of the Q partitions and P channels of the input signal
matrix X. F represents the DFT matrix dened as F = WM of size M M
and F1 as its inverse. Of course, in the nal implementation, the DFT matrix
should be substituted by much more ecient fast Fourier transform (FFT).
Being X, 2K P -dimensional (supposing 50% overlapping between the new
block and the previous one). It should be taken into account that the algorithm
adapts every K samples. W represents the lter coecient matrix adapted
in the frequency-domain (also M Q P -dimensional) while the operator
multiplies each of the elements one by one; which, in (63), represents a circular
convolution. The output vector y can be obtained as the double sum (rows)
of the Y matrix. First we obtain a M P matrix which contains the output
Adaptive Filtering Techniques for Forensic Audio 29

of each channel in the frequency-domain yp , p = 1 P , and secondly, adding

all the outputs we obtain the whole system output, y. Finally, the output in
the time-domain is obtained by using y = last Kcomponents of F1 y. Notice
that the sums are performed prior to the time-domain translation. This way
we reduce (P 1)(Q 1) FFTs in the complete ltering process. As in any
adaptive system the error can be obtained as

e = d y, (64)
with d = d(mK) d(mK + 1) d((m + 1)K 1) . The error in the
frequency-domain (for the actualization of the lter coecients) can be ob-
tained as
e = F K1 . (65)

As we can see, a block of K zeros is added to ensure a correct linear con-

volution implementation. In the same way, for the block gradient estimation,
it is necessary to employ the same error vector in the frequency-domain for
each partition q and channel p. This can be achieved by generating an error
matrix E with dimensions M Q P which contains replicas of the error
vector, dened in (65), of dimensions P and Q (E e in the notation). The
actualization of the weights is performed as

W(m + 1) = W(m) + (m)G(m). (66)

The instantaneous gradient is estimated as

G = X E. (67)

This is the unconstrained version of the algorithm which saves two FFTs
from the computational burden at the cost of decreasing the convergence
speed. The constrained version basically makes a gradient projection. The
gradient matrix is transformed into the time-domain and is transformed back
into the frequency-domain using only the rst K elements of G as
G=F . (68)

A conjugate gradient version of PBFDAF is possible by transforming the

gradient matrix to vectors and reverse [11]. The vectors g and p in (48,49)
should be changed by gl Gl , Gl = J(Wl ) and pl Pl , Pl = J(Wl Gl )
with gradient estimation obtained by averaging the instantaneous

estimates over N past values Gl = J(Wl ) = N2 k=1 Glk Wl Xlk dlk .
30 Lino Garca et. al.

3.3 Partitioned Convolution

For each input i, the AIR matrix, V, is reorganized in a column vector v =

v 1 v 2 vP of size N = LP 1 and initially partitioned in a reasonable
number Q of equally-sized blocks vq , q = 1 Q, of length K. Each of these
blocks is treated as a separate impulse response, and convolved by a standard
overlap-and-save process, using T operator (FFT windows of length L). All
input data are processed in overlapped blocks of L samples (each block at
L K samples to the last). Each block is zero-padded to length L (typically
equal to 2K), and transformed with FFT so that a collection of Q frequency-
domain lters vq is obtained. The results of the multiplications of these Q
lters vq with the FFTs of the Q input blocks are summed, producing the same
result as the unpartitioned convolution, by means of proper delays applied
to the blocks of convolved data. Finally an T1 operator (IFFT) of the rst
acummulator is made to submmit an output data block (obviosly only the last
L K block samples). Each block of input data needs to be FFT transformed
just once, and thus the number of forward FFTs is minimized [1]. The main
advantage compared to unpartitioned convolution is that the latency of the
whole ltering processing is just M points instead of 2N , and thus the I/O
delay is kept to a low value, provided that the impulse response is partitioned
in a sensible number of chunks (8-32). Figure 12 outlines the whole process.
Suposse that A = (A1 , A2 , , AQ ) is a set of multiplications of the rst
data block and B = (B1 , B2 , , BQ ) the second, then for time-index 1 it is
only necessary to consider A1 . At the next index-time, corresponding to K + 1
samples, the sum is formed with (BQ , B1 + A2 , B2 + A3 , , BQ1 + AQ ). If
C = (C1 , C2 , , CQ ) corresponds to the third block the sum is formed with
(CQ , C1 +B2 +A3 , C2 +B3 +A4 , , CQ1 +BQ ). An ecient implementation
of this sum can be implemented using a double buering technique [1].

3.4 Delayless Approach for Real-Time Applications

The ltering operation can be made delayless by operating the rst block in
the time-domain (direct convolution) while the rest of the blocks continue to
operate in the frequency domain [22]. The fast convolution starts after the
samples have been processed for direct convolution. The direct convolution
allows giving samples to the output while data is incomming. This approach
is applicable to the multirate frameworks described.

4 Adaptive Filtering Techniques

Once the theoretical foundations of the adaptive ltering have been reviewed,
the most important techniques that can be applied to forensic ltering are
Adaptive Filtering Techniques for Forensic Audio 31

input signal (subdivided in partially overlapped blocks)

blocks of L samples
1st 2nd J 1 Jth


1st spectrums Jth

v1 v2 vQ v1

seg. 22nd seg.
bloque Qth seg.
Q bloque 1st data block
1st seg. 2nd seg. Qth seg. 2nd data block

(J 1)th data block 1st seg. 2nd seg.

Jth data block 1st seg.

1 K+1 2K + 1 sum at index L

n2 bloque

T1 T1 T1 T1

1 K+1 2K + 1 output signal L

n2 bloque

Fig. 12. Partitioned convolution. Each output signal block is produced taking only
the L K last samples of the block.

4.1 Spectral Equalization

The adaptive spectral equalization is widely used for noise suppression and
corresponds to the single-input and single-output (SISO) estimator applica-
tion (class a, Fig. 5); a single microphone, P = 1, is employed. This approach
estimates a noiseprint spectra and subtracts it from the whole signal in the
The Wiener lter estimator is the result of estimating y(n) from s(n) that
minimizes the MSE y(n) s(n)2 given by y = Qx, x = s + r, and that
|x|2 |d|2
= 2 , (69)
Q = diag q1 q2 qM is a diagonal matrix which contains the spec-
tral gain in the frequency-domain; normally T is a short-time Fourier trans-
form (STFT), suitable for not stationary signals, and T1 its inverse. In this
case this algorithm is known as short-time spectral attenuation (STSA). The
M 1 vector q contains the main diagonal components of Q. d is the noise
32 Lino Garca et. al.

s(n) x(n) x y(n)

T T1


d(n) Wiener lter

Fig. 13. Spectral equalization.

spectrum (normaly unknown). In this case an estimation of the noiseprint

spectra d = r from the mixture x (noisy signal) is necessary (in intervals
when the speech is absent and only the noise is present). The STFT is dened

N mk
as xk = n=1 h(n)x(m n)WM , m = 0 M 1, where k is the time
index about which the short-time spectrum is computed, m is the discrete
frequency index, h(n) is an analysis window, N dictates the duration over
which the transform is computed, and M is a number of frequency bins at
which the STFT is computed. For stationary signals the squared-magnitude
of the STFT provides a sample estimate of the power spectrum of the under-
lying random process. This form (69) is basic to nearly all the noise reduction
methods investigated over last forty years [12]. The specic form to obtain Q
is known as the suppresion rule.

Power Subtraction

An alternative estimate from Wieners theory is achieved assuming that s can

be estimated if its magnitud is estimated as

s = |x| |d| , (70)

and the phase of the noisy signal  x can be used, if its SNR is reasonably
high, in place of  s. is an exponent and is a parameter introduced to
control the amount of noise to be subtracted ( = 1 for full subtraction and
> 1 for over subtraction). A paramount issue in spectral subtraction is to
obtain a good noise estimate; its accuracy greatly aects the noise reduction
performance [3].

4.2 Linear Prediction

The adaptive linear prediction (ALP) is employed in an attempt to separate

the deterministic y(n) s(n) and stochastic part e(n) r(n) assuming that
the noise and interference signal has a broadband spectra. ALP corresponds
Adaptive Filtering Techniques for Forensic Audio 33

to single-input and single-ouput (SISO) predictor application (class b, Fig. 5)

with a single microphone, P = 1.

s(n) x(n) z D y(n) e(n)


Fig. 14. Adaptive linear predictor.

Most signals, such as speech and music, are partially predictable and par-
tially random. The random input models the unpredictable part of the signal,
whereas the lter models the predictable structure of the signal. The aim of
linear prediction is to model the mechanism that introduces the correlation
in a signal [29]. The solution to this system corresponds to a Wiener solu-
tion (35) with the cross-correlation vector, r, slighty modied. The delay z D
in the ALP lter should be selected in such a way that d(n) = x(n) + r(n)
and d(n D) are still correlated. If D is too long, the correlation in d(n)
and d(n D) is weak and unpredictable for the ALP lter; for that reason
it cannot be canceled suitably. If D is too short, the deterministic part of
signal in d(n) and d(n D) remains correlated after D; for that reason it
can be predictable and cancelled by the ALP lter. D = 1 causes that the
voice in d(n) and d(n D) is strongly correlated. A cascade of ALP lters of
lower order independently adapted improves the modeling of the general ALP
lter. In this case, the prediction is performed in successive renements, the
adaptation steps can be greater, and thus each stage is less aected by the
disparity of eigenvalues which results in a faster convergence.

4.3 Noise Cancellation

The adaptive noise cancellation (ANC) cancels the primary unwanted noise
r(n) by introducing a canceling antinoise of equal amplitude but opposite
phase using a reference signal. This reference signal is derived from one or
more sensors located at points near the noise and interference sources where
the interest signal is weak or undetectable. A typical ANC conguration is
depicted in Fig. 15. Two microphones are used, P = 2. The primary input
d(n) = s(n) + r(n) collects the sum of unwanted noise r(n) and speech signal
s(n), and the auxiliary or reference input measures the noise signal x(n) =
ANC corresponds to multiple-input and single-output (MISO) joint-process
estimator application (class c, Fig. 5) with at least two microphones, P = 2.
34 Lino Garca et. al.


x(n) y(n) e(n)

Fig. 15. Adaptive noise cancellation.

4.4 Beamforming

Beamforming is a multiple-input and single-output (MISO) application and

consists of multichannel advanced multidimensional (space-time domain) l-
tering techniques that enhance the desired signal as well as suppress the noise
signal. In beamforming, two or more microphones are arranged in an array of
some geometric shape. A beamformer is then used to lter the sensor outputs
and amplies or attenuates the signals depending on their direction of arrival
(DOA), . The spatial response, or beampattern, of a beamformer generally fea-
tures a combination of mainlobes that may be aimed at the target sources, and
smaller sidelobes and null points aimed at the interference sources. Beampat-
terns are generally frequency-dependent, unless the beamformer is specically
designed to be frequency independent.

x1 (n)
d(n) e(n)


xP (n) AIC

b1 bP



Fig. 16. Adaptive beamforming. Robust generalized sidelobe canceller (RGSC).

Fixed beamforming (FB) allow conforming determined directivity pattern. The
adaptive block matrix (ABM) or blocking matrix, with coecient-constrained adap-
tive lters, prevents the target signal from leaking into the adaptive interference
canceller (AIC). The AIC uses norm-constrained adaptive lters that can further
improve the robustness against target signal cancellation.
Adaptive Filtering Techniques for Forensic Audio 35

The sound sources are assumed to be in the far-eld of the microphone

array, i.e. the distance of the source from the array is much greater than the
distance between the microphones (the spherical wavefronts emanating from
the sources can be approximated as plane wavefronts). Each source si (n)
arrives to microphone 1 with delay i = cos v
relative to its arrival to 2
because it has to travel an extra distance cos ; i is the DOA of si (n) and
v 355ms1 is a velocity of sound. Fvs represents the spatial samplig
interval of the waveeld; it has to fulll this inequality to avoid spatial aliasing.
The generalized sidelobe canceller (GSC) is an adaptive beamformer that
keeps track of the characteristics of the interfering signal, leading to a high
interference rejection performance. Initially, the P microphone inputs xp (n),
p = 1 P , go through the FB that directs the beam towards the expected
DOA. The beamformer output y(n) = h, x(n) contains the enhanced sig-
nal originating from the pointed direction, which is used as a reference by
the ABM. The coecient vector h has to fulll both spatial and temporal
constrains Ch = c, h = C[CH C]1 c. The ABM adaptively subtracts the
signal of interest, represented by the reference signal y(n), from each channel
input xp (n), and provides the interference signals. The columns of C must be
pairwise orthogonal to the columns of the blocking matrix B, CH B = 0. The
quiescent vector h is a component independently of data and w = h Bg is
a lter that satises the linear constrains CH w = CH (h Bg) = CH h = c.
The upper signal path in Fig. 16 has to be orthogonal to the lower signal path.
In order to suppress only those signals that originate from a specic track-
ing region, the adaptive lter coecients are constrained within predened
boundaries [3]. These boundaries are specied based on the maximum allowed
deviation between the expected DOA and the actual DOA. The interference
signals, obtained from the ABM, are passed to the AIC, which adaptively
removes the signal components that are correlated to the interference signals
from the beamformer output y(n). The norm of the lter coecients in the
AIC is constrained to prevent them from growing excessively large. This min-
imizes undesirable target signal cancellation, when the target signal leaks into
the AIC, further improving the robustness of the system [31].
In noise reduction systems, the beamformer can be used to either reject
the noise (interference) by attenuating signals from certain DOAs, or focus
on the desired signal (target) by amplifying signals from the target DOA and
attenuating all signals that are not from the target DOAs. For non real-time
audio forensic applications it is possible to select a set of DOAs to be tested.
Therefore adaptive algorithms with directional constrains, like a RGSC, can
be exploited to achieve better noise reduction performance.

4.5 Deconvolution

Both blind signal separation (BSS), also known as blind source separation,
and multichannel blind deconvolution (MBD) problems are a type of inverse
problems with similarities and subtle dierences between them: in the MBD
36 Lino Garca et. al.

only one source is considered, and thus the system is single-input single-output
(SISO), while in BSS there are always multiple independent sources and the
mixing system is MIMO; the interest of MBD is to deconvolve the source from
the AIRs, while the task of BSS is double: on the one hand the sources must
be separated, on the other hand the sources must be deconvolved from the
multiple AIRs since each sensor collects a combination of every original source
convolved by diferent lters (AIRs) according to (2) [13].

z D d(n) = si (n)

si (n) x1 (n) y(n) e(n)

v1 w1

xP (n)
vP wP


Fig. 17. Multichannel blind deconvolution.

x1 (n) y(n)
v1 w1
si (n)

xP (n)
vP wP


Fig. 18. Blind source factor separation.

In both cases, the blind deconvolution or equalization approach as well as

the blind separation one, must estimate adaptively the inverse of the convolu-
tive system that allows recovering the input signals and suppressing the noise.
The goal is to adjust W so that WV = PD, where P is a permutation matrix
and D is a diagonal matrix whose (p, p)th is p z p ; p is a nonzero scalar
weigthing, and p is an integer delay.
BSS deals with the problem of separating I unknown sources by observ-
ing P microphone signals. In the underdetermined case (P < I) there are
innitely possible vectors s(n) that satisfy (3). There are mainly two ways
Adaptive Filtering Techniques for Forensic Audio 37

to achieve the minimum norm solution. In the rst, the right generalized in-
verse of V is estimated and then applied to the set of microphone signals
x(n). Another class of algorithms employ the sparseness of speech signal to
design better inversion strategies and identify the minimum norm solution.
Many techniques of convolutive BSS have been developed by extending meth-
ods originally designed for blind deconvolution of just one channel. A usual
practice is to use blind source factor separation (BSFS)Blind source factor sep-
aration technique, where one source (factor) is separated from the mixtures,
and combine it with a deationary approach, where the sources are extracted
one by one after deating, i.e. removing, them from the mixed signals. The
MIMO FIR lter W used for BSS becomes a multiple-input single-output
(MISO) depicted in Fig. 18. The output y(n) corresponds to (11) and the
tap-stacked column vector containing all demixing lter weights dened in
(7) is obtained as

u = Rp, (71)
w= ,
u Ru
where R is a block matrix where its blocks are the correlation matrices Rpq be-
tween the p-th channel and q-th channel dened in (36) and p is a block vector
where its blocks are the cross-cumulant vector p = cum{x(n), y(n) y(n)}
[13]. The second step in (71) is just the normalization of the output signal
y(n). This is apparent left multiplying by x(n).
The deationary BSS algorithm for i = 1 I sources can be summa-
rized as following: one source is extracted with the BSFS iterative scheme
till convergence (71) and the ltering of the microphone signals with the
estimated lters from the BSFS method (11) is performed; the contribu-
tion of the extracted source into the mixtures xp , p = 1 P , is estimated
(with the LS criterion) and the contribution of the o-th source into i-th
mixture is computed by using the estimated
lter b, c(n) = b, y(n) with
y(n) = y(n) y(n 1) y(n B + 1) , B << L; deate the contribution
c(n) from the p-th mixture, xp (n) = xp (n) c(n), p = 1 P . This method is
very suitable for audio forensic application where only one source should be
extracted, i.e. speech.
It is possible to consider the deationay BSFS (DBSFS) structure as a
GSC. ABM exactly corresponds to the deating lters of the deationary ap-
proach. By comparing the dierent parts, i.e. the BSFS block and the xed
beamformer, it is concluded that it may be possible to construct similar algo-
rithms to those of GSC.

5 Conclusions
This chapter is an advanced tutorial about multichannel adaptive ltering for
forensic audio. Dierent techniques have been examined in a common foun-
38 Lino Garca et. al.

dation. Several approaches of forensic ltering were presented as the number

of channels increases.
The spectral equalization (power subtraction), in general, can achieve more
noise reduction than an ANC and a beamformer method. However, this is
based on the noise spectrum estimator instead of the unknown noise spectra
at each time producing a distortion known as musical noise (because of the
way it sounds). The performance of ANC depends on the coherence between
the input noisy signal and the reference noise signal. Only if the coherence is
very high the results are spectacular, therefore, this fact limits its application
to particular cases. The amount of noise that can be canceled by a beamformer
relies on the number of microphones in the array and on the SNR of the
input signal. More microphones can lead to more noise reduction. However,
the eectiveness of a beamformer in suppressing directional noise depends
on the angular separation between signal and the noise source [3]. The ALP
method is very simple because only second order statistics are required, but
the estimation is only optimal if the residue is i.i.d. Gaussian [27].
All these techniques are narrowly connected. The linear prediction of x(n)
is nothing but the deconvolution of x(n) [27]. In [28], the problem of Wiener
system blind inversion using source separation methods is addressed. This
approach can also be used for blind linear deconvolution. In [13] the link
between the deationary approach (the extension of the single channel blind
deconvolution algorithm) and the traditional GSC structure is showed. Several
strategies between dierent approaches are also possible, i.e. in [29], a Wiener
lter, that uses linear prediction to estimate the signal spectrum, is presented.
The best lter to enhance a particular recording will be chosen based
on experience and experimentation [20]. Nevertheless, the algorithm devel-
oper would nd it useful to have a quality measure that helps to compare,
in general terms, the performance of dierent implementations of a certain
algorithm [30]. One substantial ingredient of this performance is the intelli-
gibility attained after processing the recording, or even better the increase
of intelligibility compared to the unprocessed sample. Therefore, one possible
way to measure the performance of an enhancement algorithm, and probably
the best, would be to use a panel of listeners and one of the subjective tests
introduced in Sect. 1.3. To attain signicant results, dierent speech record-
ings with dierent types and degrees of noise and distortion should be used
as inputs to the algorithm, and therefore the task would probably become
unapproachable in terms of time and eort, setting aside the fact that the
experiment would hardly be repeatable.
For these reasons, an objective intelligibility measure would be of great
help to compare, or even adjust, voice enhancement algorithms. The STI mea-
sure, introduced in Sect. 1.3, is probably the most appropriate one, because
it can deal with more types of degradations than other objective intelligibil-
ity measures. As previously mentioned, to compute the STI of a transmis-
sion system, that in our case comprises the acoustic environment where the
recorded was performed, followed by the enhancement system, either the im-
Adaptive Filtering Techniques for Forensic Audio 39

pulse response of the whole system or a processed version of an special probe

signal is needed. As adaptive lters are time variant, they can not be identied
by an impulse response, thus the use of the probe signal is mandatory.
In order to properly monitor the performance of the algorithms, dierent
types and degrees of degradations should be imposed to the test signal. The
model used to deal with degradations can be as simple as an additive noise,
for a mono version of the test signal corrupted by random noise or a second
talker speech, or as complex as a virtual room simulator for early reexions
and a stocastic reverberation generator, for a detailed acoustic model of the
recording room, where several noise sources can be placed in dierent places.
Measured impulse responses of a real chamber is another option to obtain
very realistic mono or multi-channel virtual recordings.
Although STI seems useful for this kind of analysis, providing a practical
systematic tool to measure the degree of enhancement attained, it should
be kept in mind that it was not designed to test adaptive lters, and its
behaviour during the adaptation periods of the algorithms may be misleading
[15]. Hence, the obtained results should be carefully contrasted with subjective
listening tests. An example of the use of SRT to evaluate the performance of
beamforming techniques can be found in [4].

1. Armelloni E, Giottoli C, Farina A (2003) Implementation of Real-time Parti-
tioned Convolution on a DSP Board, 2003 IEEE Workshop on the Applications
of Signal Processing to Audio and Acoustics, page(s):7174.
2. Barnett, P.W, Knight, R.D (1995) The Common Intelligibility Scale, Proc.
I.O.A. 17(7):199-204.
3. Benesty J, Huang Y (2003) Adaptive Signal Processing (Applications to Real-
World Problems). Springer, Berlin Heidelberg New York Hong Kong London
Milan Paris Tokyo
4. Beracoechea, J (2007) Codicacin de Audio Multicanal para entornos de tipo
Ventana Acstica, Ph D. Thesis, Universidad Politcnica de Madrid
5. Betts D, French A, Hicks C, Reid G (2005) The Role of Adaptive Filtering in
Audio Surveillance. AES 26th International Conference.
6. Boray G.K, Srinath M.D (1992) Conjugate Gradient Techniques for Adaptive
Filtering. IEEE Transactions on Circuits and Systems 39(1):1-10.
7. Chau E.Y-H (2001) Adaptive Noise Reduction Using A Cascaded Hybrid Neu-
ral Network. MS Thesis, University of Guelph, Ontario
8. Crochiere R.E, Rabiner L.R (1983) Multirate Digital Signal Processing.
Prentice-Hall, London Sidney Toronto Mexico New Delhi Tokyo Singapore Rio
de Janeiro
9. French, N.R, Steinberg, J.C (1947) Factors governing the intelligibility of speech
sounds, The Journal of the Acoustical Society of America 19:90119.
10. Friedlander B (1982) Lattice Filters for Adaptive Processing. Proceedings of
the IEEE 70(8):829867.
40 Lino Garca et. al.

11. Garca L (2006) Cancelacion de Ecos Multicanal. PhD Thesis, Universidad

Politecnica de Madrid, Spain
12. Gay S.L, Benesty J (2000) Acoustic Signal Processing for Telecommunication.
Kluwer Academic Publishers, Boston Dordrecht London
13. Gkalelis N (2004) Undetermined Blind Source Separation for Speech Signals.
MS Thesis, Friedrich-Alexander-Universitat Erlangen-Nurnberg, Germany
14. Glentis G.-O, Berberidis K, Theodoridis S (1999) Ecient least square adaptive
algorithms for FIR transversal ltering: A unied view, IEEE Signal Processing
Magazine 16(4):1341
15. Goldsworthy R.L, Greenberg J.E (2004) Analysis of speech-based speech trans-
mission index methods with implications for nonlinear operations, The Journal
of the Acoustical Society of America 116(6):36793689.
16. Haykin S (2002) Adaptive Filter Theory. Prentice-Hall, Inc., New Jersey
17. Honig M.L, Messerschmitt D.G (1984) Adaptive Filters: Structures, Algorithms
and Applications. Kluwer Academic Publishers, Boston The Hague London
18. IEC Standard 60268-16 (2003)
19. ISO standard 9921 (2003)
20. Koenig B.E, Lacey D.S, Killion S.A (2007) Forensic enhancement of digital
audio recordings, Journal of de Audio Engineering Society, 55(5):352371.
21. Mayyas K (2002) Stereophonic Acoustic Echo Cancellation Using Lattice
Orthogonalization, IEEE Transactions on Speech and Audio Processing,
22. Morgan D.R, Thi J.C (1995) A Delayless Subband Adaptive Filter Architecture,
IEEE Transactions on Signal Processing, 43(8):18191830.
23. Paez Borrallo J.M, Otero M.G (1992) On The Implementation of a Partitioned
Block Frequency Domain Adaptive Filter (PBFDAF) For Long Acoustic Echo
Cancellation, Signal Processing, 27(3)301315.
24. Plomp R., Mimpen, A.M (1979) Improving the reliability of testing the speech
reception threshold for sentences, Audiology 8:4352.
25. Reilly J.P, Wilbur M, Seibert M, Ahmadvand N (2002) The Complex Subband
Decomposition and its Application to the Decimation of Large Adaptive Fil-
tering Problems, IEEE Transactions on Signal Processing, 50(11):27302743.
26. Shynk J.J (1992) Frequency-domain and Multirate Adaptive Filtering, IEEE
Signal Processing Magazine, 9(1):1437.
27. Sole-Casals J, Jutten C, Taleb A (2000) Source Separation Techniques Applied
to Linear Prediction, 2th International Workshop on Independent Component
Analysis and Blind Source Separation Proceedings ICA2000, page(s):193198.
28. Taleb A, Sole-Casals J, Jutten C (1999) Blind Inversion of Wiener Systems,
IWANN 99, page(s):655664.
29. Vaseghi S.V (1996) Advanced Signal Processing and Digital Noise Reduction.
John Willey & Sons Ltd. and B.G. Teubner, Chichester New York Brisbane
Toronto Singapore Stuttgart Leipzig
30. Yi H, Philipos C.L (2007) A comparative intelligibility study of single-
microphone noise reduction algorithms, The Journal of the Acoustical Society
of America, 122(3):1777-1786.
31. Yoon B-Y, Tashev I, Acero A (2007) Robust Adaptive Beamforming Algorithm
Using Instantaneous Direction Of Arrival With Enhanced Noise Suppression
Capability. IEEE International Conference on Acoustics, Speech and Signal
Processing 1:I-133I-136.

Acoustic impulse response, 3, 6, 25, 30, Forensic ltering, 2, 3, 16, 30, 38

Adaptive ltering, 1, 9, 10, 25 Joint-process estimation, 11, 12, 14, 15,
Adaptive linear prediction, 32 18, 33
Adaptive noise cancellation, 33
Adaptive spectral equalization, 31, 38 Lattice-ladder joint-process estimation,
Audio enhancement, 1, 8 13, 14, 23
Audio restoration, 8
Multichannel adaptive ltering, 2, 12,
Beamforming, 34 14, 37
Blind deconvolution, 35, 36
Blind source separation, 35 Noise, 26, 810, 3134, 36, 38
Noise reduction, 2, 4, 8, 9, 31, 35, 38
Cost functions, 16
Speech intelligibility, 8, 9, 38
Forensic audio, 13, 5, 8, 37 Speech listenability, 8