Low-artifact source separation using probabilistic latent component analysis

Nasser Mohammadiha, Paris Smaragdis, Arne Leijon

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Abstract

We propose a method based on the probabilistic latent component analysis (PLCA) in which we use exponential distributions as priors to decrease the activity level of a given basis vector. A straightforward application of this method is when we try to extract a desired source from a mixture with low artifacts. For this purpose, we propose a maximum a posteriori (MAP) approach to identify the common basis vectors between two sources. A low-artifact estimate can now be obtained by using a constraint such that the common basis vectors in the interfering signal's dictionary tend to remain inactive. We discuss applications of this method in source separation with similar-gender speakers and in enhancing a speech signal that is contaminated with babble noise. Our simulations show that the proposed method not only reduces the artifacts but also increases the overall quality of the estimated signal.

Original languageEnglish (US)
Title of host publication2013 IEEE Workshop on Applications of Signal Processing to Audio and Acoustics, WASPAA 2013
DOIs
StatePublished - 2013
Event2013 14th IEEE Workshop on Applications of Signal Processing to Audio and Acoustics, WASPAA 2013 - New Paltz, NY, United States
Duration: Oct 20 2013Oct 23 2013

Publication series

NameIEEE Workshop on Applications of Signal Processing to Audio and Acoustics

Other

Other2013 14th IEEE Workshop on Applications of Signal Processing to Audio and Acoustics, WASPAA 2013
CountryUnited States
CityNew Paltz, NY
Period10/20/1310/23/13

Keywords

  • Artifact Reduction
  • Dictionary Learning
  • Nonnegative Matrix Factorization (NMF)
  • PLCA
  • Source Separation

ASJC Scopus subject areas

  • Electrical and Electronic Engineering
  • Computer Science Applications

Fingerprint Dive into the research topics of 'Low-artifact source separation using probabilistic latent component analysis'. Together they form a unique fingerprint.

Cite this