Discriminative NMF and its application to single-channel source separation

Felix Weninger, Jonathan Le Roux, John R. Hershey, Shinji Watanabe

Research output: Contribution to journalArticle

63 Citations (Scopus)

Abstract

The objective of single-channel source separation is to accurately recover source signals from mixtures. Non-negative matrix factorization (NMF) is a popular approach for this task, yet previous NMF approaches have not optimized directly this objective, despite some efforts in this direction. Our paper introduces discriminative training of the NMF basis functions such that, given the coefficients obtained on a mixture, a desired source is optimally recovered. We approach this optimization by generalizing the model to have separate analysis and reconstruction basis functions. This generalization frees us to optimize reconstruction objectives that incorporate the filtering step and SNR performance criteria. A novel multiplicative update algorithm is presented for the optimization of the reconstruction basis functions according to the proposed discriminative objective functions. Results on the 2nd CHiME Speech Separation and Recognition Challenge task indicate significant gains in source-to-distortion ratio with respect to sparse NMF, exemplar-based NMF, as well as a previously proposed discriminative NMF criterion.

Original languageEnglish
Pages (from-to)865-869
Number of pages5
JournalUnknown Journal
Publication statusPublished - 2014
Externally publishedYes

Fingerprint

Source separation
Source Separation
Matrix Factorization
Factorization
factorization
Basis Functions
matrices
Discriminative Training
Non-negative Matrix Factorization
Optimization
optimization
Multiplicative
Filtering
Objective function
Update
Optimise
education
Coefficient
coefficients

ASJC Scopus subject areas

  • Language and Linguistics
  • Human-Computer Interaction
  • Signal Processing
  • Software
  • Modelling and Simulation

Cite this

Discriminative NMF and its application to single-channel source separation. / Weninger, Felix; Le Roux, Jonathan; Hershey, John R.; Watanabe, Shinji.

In: Unknown Journal, 2014, p. 865-869.

Research output: Contribution to journalArticle

Weninger, Felix ; Le Roux, Jonathan ; Hershey, John R. ; Watanabe, Shinji. / Discriminative NMF and its application to single-channel source separation. In: Unknown Journal. 2014 ; pp. 865-869.
@article{2b78e919d51a42b1821db7bfa8c42a9c,
title = "Discriminative NMF and its application to single-channel source separation",
abstract = "The objective of single-channel source separation is to accurately recover source signals from mixtures. Non-negative matrix factorization (NMF) is a popular approach for this task, yet previous NMF approaches have not optimized directly this objective, despite some efforts in this direction. Our paper introduces discriminative training of the NMF basis functions such that, given the coefficients obtained on a mixture, a desired source is optimally recovered. We approach this optimization by generalizing the model to have separate analysis and reconstruction basis functions. This generalization frees us to optimize reconstruction objectives that incorporate the filtering step and SNR performance criteria. A novel multiplicative update algorithm is presented for the optimization of the reconstruction basis functions according to the proposed discriminative objective functions. Results on the 2nd CHiME Speech Separation and Recognition Challenge task indicate significant gains in source-to-distortion ratio with respect to sparse NMF, exemplar-based NMF, as well as a previously proposed discriminative NMF criterion.",
author = "Felix Weninger and {Le Roux}, Jonathan and Hershey, {John R.} and Shinji Watanabe",
year = "2014",
language = "English",
pages = "865--869",
journal = "Nuclear Physics A",
issn = "0375-9474",
publisher = "Elsevier",

}

TY - JOUR

T1 - Discriminative NMF and its application to single-channel source separation

AU - Weninger, Felix

AU - Le Roux, Jonathan

AU - Hershey, John R.

AU - Watanabe, Shinji

PY - 2014

Y1 - 2014

N2 - The objective of single-channel source separation is to accurately recover source signals from mixtures. Non-negative matrix factorization (NMF) is a popular approach for this task, yet previous NMF approaches have not optimized directly this objective, despite some efforts in this direction. Our paper introduces discriminative training of the NMF basis functions such that, given the coefficients obtained on a mixture, a desired source is optimally recovered. We approach this optimization by generalizing the model to have separate analysis and reconstruction basis functions. This generalization frees us to optimize reconstruction objectives that incorporate the filtering step and SNR performance criteria. A novel multiplicative update algorithm is presented for the optimization of the reconstruction basis functions according to the proposed discriminative objective functions. Results on the 2nd CHiME Speech Separation and Recognition Challenge task indicate significant gains in source-to-distortion ratio with respect to sparse NMF, exemplar-based NMF, as well as a previously proposed discriminative NMF criterion.

AB - The objective of single-channel source separation is to accurately recover source signals from mixtures. Non-negative matrix factorization (NMF) is a popular approach for this task, yet previous NMF approaches have not optimized directly this objective, despite some efforts in this direction. Our paper introduces discriminative training of the NMF basis functions such that, given the coefficients obtained on a mixture, a desired source is optimally recovered. We approach this optimization by generalizing the model to have separate analysis and reconstruction basis functions. This generalization frees us to optimize reconstruction objectives that incorporate the filtering step and SNR performance criteria. A novel multiplicative update algorithm is presented for the optimization of the reconstruction basis functions according to the proposed discriminative objective functions. Results on the 2nd CHiME Speech Separation and Recognition Challenge task indicate significant gains in source-to-distortion ratio with respect to sparse NMF, exemplar-based NMF, as well as a previously proposed discriminative NMF criterion.

UR - http://www.scopus.com/inward/record.url?scp=84910065215&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=84910065215&partnerID=8YFLogxK

M3 - Article

AN - SCOPUS:84910065215

SP - 865

EP - 869

JO - Nuclear Physics A

JF - Nuclear Physics A

SN - 0375-9474

ER -