BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//talks.cam.ac.uk//v3//EN
BEGIN:VTIMEZONE
TZID:Europe/London
BEGIN:DAYLIGHT
TZOFFSETFROM:+0000
TZOFFSETTO:+0100
TZNAME:BST
DTSTART:19700329T010000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=-1SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:+0100
TZOFFSETTO:+0000
TZNAME:GMT
DTSTART:19701025T020000
RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=-1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
CATEGORIES:Signal Processing and Communications Lab Seminars
SUMMARY:Itakura-Saito nonnegative factorizations of the po
wer spectrogram for music signal decomposition - D
r Cedric Fevotte\, CNRS - TELECOM ParisTech
DTSTART;TZID=Europe/London:20100318T141500
DTEND;TZID=Europe/London:20100318T151500
UID:TALK22370AThttp://talks.cam.ac.uk
URL:http://talks.cam.ac.uk/talk/index/22370
DESCRIPTION:Nonnegative matrix factorization (NMF) is a popula
r linear regression technique in the fields of mac
hine learning and signal/image processing. \nMuch
research about this topic has been driven by appli
cations in audio. \nNMF has been for example appli
ed with success to automatic music transcription a
nd audio source separation\, where the data is usu
ally taken as the magnitude spectrogram of the sou
nd signal\, and the Euclidean distance or Kullback
-Leibler divergence are used as measures of fit be
tween the original spectrogram and its approximate
factorization.\n\nAfter a brief overview of NMF\,
in this presentation we will show evidence of the
relevance of considering factorization of the pow
er spectrogram\, with the Itakura-Saito (IS) diver
gence. Indeed\, IS-NMF is shown to be connected to
maximum likelihood inference of variance paramete
rs in a well-defined statistical model of superimp
osed Gaussian components and this model is in turn
shown to be well suited to audio. \nFurthermore\,
the statistical setting opens doors to Bayesian a
pproaches and to a variety of computational infere
nce techniques. We discuss in particular model ord
er selection strategies and Markov regularization
of the activation matrix\, to account for time-per
sistence in audio.\n\nThis presentation will also
adress extensions of NMF to the multichannel case\
, in both instantaneous or convolutive recordings\
, possibly underdetermined\, leading to nonnegativ
e tensor factorizations under novel structures. We
will present in particular audio source separatio
n results of real-world stereo musical excerpts.\n
\nReferences :\n\nC. Févotte\, N. Bertin and J.-L.
Durrieu. "Nonnegative matrix factorization with t
he Itakura-Saito divergence. With application to m
usic analysis\," Neural Computation\, vol. 21\, no
3\, Mar. 2009 http://www.tsi.enst.fr/~fevotte/Jou
rnals/neco09_is-nmf.pdf\n\nA. Ozerov and C. Févott
e. "Multichannel nonnegative matrix factorization
in convolutive mixtures for audio source separatio
n\," IEEE Trans. Audio\, Speech and Language Proce
ssing\, 2010 (to appear) http://www.tsi.enst.fr/~f
evotte/TechRep/techrep09_multinmf.pdf\n
LOCATION:LR5\, Engineering\, Department of
CONTACT:Rachel Fogg
END:VEVENT
END:VCALENDAR