Page 22 - EURECOM - RA2011GB

Basic HTML Version

Focus on…
An illustration of acoustic echo cancellation and noise
compensation through adaptive filtering and postfiltering.
INTEL et EURECOM,
partners in acoustic echo cancellation technology
T
oday, we expect to use our mobile tel-
ephones on themove and in almost any
situation. This can be in the office, the
car, or perhaps a busy café. The level of noise in
eachof theseenvironmentscanvarygreatlyand
sometimes be high enough to disturb a remote
listener. Listeningqualitymayalsobedegraded
by the couplingof speechbetweenamobile ter-
minal loudspeakerandmicrophone.Undersuch
circumstancesa remotespeaker cansometimes
hear an echo of their own voice and, when the
network delay is significant, then echo can also
contribute to listener fatigue.
Researchers and the mobile telecommuni-
cations industry have consequently invested
significantly to develop noise compensation
algorithms to reduce the level of perceived noise
whenamobiletelephoneisusedinadverse,noisy
environments.Acousticechocancellationtechnol-
ogyhasalsobeendevelopedtoattenuatethelevel
of echo in an uplink signal and thus to improve
communicationquality.Today’snoisecompensa-
tion algorithms reach their limits in particularly
high levels of noise, however, whereas current
acoustic echo cancellation technology tends not
tocopewell forparticularlysmallmobileterminals
where the acoustic echo path canbe non-linear.
Over the last three years EURECOMhas devel-
opedapartnershipwithIntelMobileCommunica-
tions (formerlyInfineonTechnologies) todevelop
newapproaches tonoisecompensationandnon-
linear acoustic echo cancellation. The earliest
of this work has resulted in new cascaded and
non-linearmodelsofmobileterminal loudspeak-
ers, amplifiers and acoustic channels. Nearing
its completion, we will soon see this technology
finding itsway intothenextgenerationofmobile
telephones.
contact:
Nicholas.Evans@eurecom.fr
media documents that frameworks such as the
one we have proposed for learning andmining
multimedia semantic concept fromweb
sources [1] can performmore efficiently.
W 
ith over 48 hours of video
uploaded everyminute on
YouTube alone, the need for
efficient and scalablemultimediamining
and retrieval approaches as never been
stronger. The volume of data coupledwith the
number of users is creating new challenges
for multimedia researchers in terms of
algorithmic scalability, effectiveness and
efficiency. Fortunately, a significant amount
of shared, onlinemedia is tagged, either
manually by the owner, or automatically
withmetadata originating from the capturing
device (i.e. geolocalization, EXIF, etc…). Can
such unprecedentedwealth of data play a
role in solving the long-standingmultimedia
semantic gap challenge? This is a research
directionwe have recently been studying and
exploring.
We address the problemof video annotation
using both content-based information
originating fromvisual characteristics and
textual information (metadata and user tags)
associatedwith themultimedia documents.
Traditionally, the visual model corresponding
to a concept (or label) is learnt usingmachine
learning approaches (such as Support Vector
WebScaleMultimedia Collections,
Mining andRetrieval
Machines) on a large dataset for which human
experts have provided annotations. However,
annotation efforts for a large dataset are
both error prone and time consuming. Media
uploaded on social sharing platforms are
commonly accompaniedwithmetadata such
as tags, description, etc..., which are provided
“for free” by users themselves. Although
this information is usually sparse and often
inaccurate, there are somany such shared
contact:
Benoit.Huet@eurecom.fr
web:
http://www.eurecom.fr/fr/people/huet-benoit/research
Media
+
Keywords
Mulimedia
ConceptsModels
Categorized
Samples
Semantic Tree
Automatically
Labeled/Indexed
Videos
Deep
Learning/Training
J
C Z R
Postfilter
AEC
Loudspeaker-enclosure-microphone (LEM) system
NEAR
END
FAR
END
noise
speech
echo
ê
(
k
)
_
e
(
k
)
s
(
k
)
x
(
k
)
n
(
k
)
+
y
(
k
)
s
(
k
)
Asaptive
filter
22
EURECOM
Graduate school and research center in communication systems