Postdoc position at IRISA, Rennes, France

[Emmanuel Vincent <emmanuel.vincent@xxxxxxxx>]

Dear list,

We are seeking to recruit a postdoctoral researcher on adaptive spectral 
modeling of audio signals, applied to source separation and object-based 
sound scene description (full subject below). The successful candidate 
will work in the METISS group at the public research institute IRISA 
under the supervision of Drs. Emmanuel Vincent and Rémi Gribonval.

Prospective candidates should have a background in signal processing and 
hold a PhD from may 2007 or after or be about to obtain one. Additional 
knowledge about audio is appreciated. Informal enquiries may be made to 
Emmanuel Vincent (emmanuel.vincent@xxxxxxxx).

This appointment will start fall 2008 and salary will be at 28000 euros 
per annum. Applications must be submitted online before february 15th at
http://www.inria.fr/travailler/mrted/en/postdoc/details.html?nPostingTargetID=4835



Semi-adaptive spectrum models for audio signal processing

Most audio signals are mixtures of several sources present at the same 
time: speakers, musical instruments, natural sounds. The modeling of 
these sources is the core problem behind many audio signal processing 
tasks, such as source separation, content classification and 
speech/music transcription. The most general family of models consists 
of representing the short-term power spectrum of each source as a linear 
combination of basis spectra learned on single-source training data. The 
adaptation of these models to the considered mixture is bound to improve 
performance. However existing adaptation techniques often fail, due to 
the large number of free parameters per source.

Our team has recently started to investigate a new adaptation paradigm 
for such models, whereby generic constraints satisfied by a range of 
audio sources are specified on the model parameters. This reduces the 
number of free parameters, hence improving the quality of adaptation and 
removing the need for single-source training data. For instance, 
harmonicity can be enforced by representing each basis spectrum as a 
linear combination of fixed spectra each representing a few adjacent 
harmonic partials. The spectral envelope of each basis spectrum is then 
learned by adapting the combination weights.

The first goal of this postdoc is to validate this paradigm in a fully 
general context by designing and testing a larger set of appropriate 
constraints. Possible constraints include: inharmonicity or wideband 
character of the basis spectra, source-filter model of the spectral 
envelope, transient or continuous character of the combination weight 
sequences. A second goal is to propose a way of exploiting the available 
spatial information for model adaptation. A promising approach consists 
of modeling and tracking the interchannel intensity difference of each 
source, in addition to its power spectrum.

The proposed paradigm is expected to have a high impact on the 
processing of audio data within large databases, where single-source 
training data are typically unavailable due to the huge range of 
possible sources. The results will be primarily evaluated for the task 
of source separation on a large range of audio mixtures, including 
speech, music and natural sound scenes. Depending on the research 
background of the applicant, additional tasks will be considered such as 
temporal decomposition of speech, multiple pitch estimation of music or 
sound object identification within natural sound scenes.


--
Emmanuel Vincent
METISS Project-Team
IRISA-INRIA
Campus de Beaulieu, 35042 Rennes cedex, France
Phone: +332 9984 2269 - Fax: +332 9984 7171
Web: http://www.irisa.fr/metiss/members/evincent/