Subject: Postdoctoral position in real-time music source separation From: Emmanuel Vincent <emmanuel.vincent@xxxxxxxx> Date: Tue, 7 Sep 2010 17:24:47 +0200 List-Archive:<http://lists.mcgill.ca/scripts/wa.exe?LIST=AUDITORY>The METISS team at INRIA Rennes, France, is offering a postdoc position in real-time music source separation in the context of a project with Audionamix, the leading company in source separation (see details below). Applications including a full resume, a letter of motivation and up to three reference letters must be sent by email to the principal investigator before October 1, 2010. Phone interviews of selected candidates will be held during the first week of October. TITLE: Real-time music source separation DURATION: 2.5 years RECRUITMENT DATE: as soon as possible and no later than January 1, 2011 SALARY: according to experience PRINCIPAL INVESTIGATOR: Emmanuel Vincent (emmanuel.vincent@xxxxxxxx) CO-PRINCIPAL INVESTIGATOR: Rémi Gribonval (remi.gribonval@xxxxxxxx) DESCRIPTION OF THE PROJECT: As a consequence of the ubiquity of 3D audio devices, music professionals and listeners are expecting increasingly advanced stereo-to-3D playback systems. While current systems rely on simple spatial filtering algorithms, real-time access to individual sound sources is necessary for improved spatialization and interaction. Today's audio source separation algorithms cannot be used in this context since they typically operate offline by learning source models over the full signal duration. This postdoctoral project aims to design real-time music source separation algorithms. This work will be based on the state-of-the-art "variance modeling" paradigm [1] making it possible to combine alternative source models such as GMM and NMF currently at the core of industrial source separation systems. Three challenges will be investigated in particular: - determining the best combination of models for the extraction of a target class of sources given computation and parallelization constraints, - proposing a general procedure for robust online learning of the model parameters from small amounts of data, - estimating advanced spatial parameters in addition to the source directions of arrival, e.g. their distance to the microphones or the room reverberation time, and exploiting them for dereverberation Promising research tracks can be found in the literature about online GMM or NMF learning [2,3] and variance-based reverberation modeling [4]. [1] A. Ozerov, E. Vincent and F. Bimbot, "A general modular framework for audio source separation", in Proc. 9th Int. Conf. on Latent Variable Analysis and Signal Separation (LVA/ICA), 2010. [2] Y. Zhang and M.S. Scordilis, "Effective online unsupervised adaptation of Gaussian mixture models and its application to speech classification", Pattern Recognition Letters 29(6), 2008. [3] B. Cao, D. Shen, J.-T. Sun, X. Wang, Q. Yang and Z. Chen, "Latent factor detection and tracking with online non negative matrix factorization", in Proc. International Joint Conferences on Artificial Intelligence (IJCAI), 2007. [4] N.Q.K. Duong, E. Vincent and R. Gribonval, "Under-determined reverberant audio source separation using a full-rank spatial covariance model", IEEE Transactions on Audio, Speech and Language Processing 18(7), 2010. WORK ENVIRONMENT: INRIA, the French National Institute for Research in Computer Science and Control plays a leading role in the development of Information Science and Technology (IST) in Europe. The METISS team at INRIA Rennes gathers a staff of 20 people focusing on audio signal processing research. This position is part of the i3DMusic project supported by EUREKA aiming to interactively respatialize mono or stereo music content in real time. It will involve regular exchanges and collaboration with the project coordinator Audionamix (http://www.audionamix.com/) in Paris and the other partners Sonic Emotion (http://www.sonicemotion.com/) and EPFL in Switzerland. CANDIDATE PROFILE: Prospective candidates must hold or be about to defend a PhD in audio signal processing. Proficient coding in Matlab or C++ is necessary. Additional knowledge about musical audio, 3D audio rendering or parallel computing would be an asset. -- Emmanuel Vincent METISS Project-Team INRIA Rennes - Bretagne Atlantique Campus de Beaulieu, 35042 Rennes cedex, France Phone: +332 9984 2269 - Fax: +332 9984 7171 Web: http://www.irisa.fr/metiss/members/evincent/