Reverberation-robust underdetermined source separation with non-negative tensor double deconvolution. Murata, N., Kameoka, H., Kinoshita, K., Araki, S., Nakatani, T., Koyama, S., & Saruwatari, H. In 2016 24th European Signal Processing Conference (EUSIPCO), pages 1648-1652, Aug, 2016.
Reverberation-robust underdetermined source separation with non-negative tensor double deconvolution [pdf]Paper  doi  abstract   bibtex   
Source separation using an ad hoc microphone array can be useful for enhancing speech in such applications as teleconference systems without the need to prepare special devices. However, the positions of the sources (and the microphones when using an ad hoc microphone array) can change during recording, thus violating the commonly made assumption in many source separation algorithms that the mixing system is time-invariant. This paper proposes an extension of the multichannel nonnegative matrix factorization (NMF) approach to deal with the problem of underdetermined source separation in time-variant reverberant environments. The proposed method models the mixing system as a non-negative convolutive mixture based on the concept of a “semi-time-variant system” to handle the reverberation in a room as well allowing for relatively small changes in the source/microphone positions. It also models the power spectrogram of each sound source using the convolutive NMF model to consider the local dynamics of speech.
@InProceedings{7760528,
  author = {N. Murata and H. Kameoka and K. Kinoshita and S. Araki and T. Nakatani and S. Koyama and H. Saruwatari},
  booktitle = {2016 24th European Signal Processing Conference (EUSIPCO)},
  title = {Reverberation-robust underdetermined source separation with non-negative tensor double deconvolution},
  year = {2016},
  pages = {1648-1652},
  abstract = {Source separation using an ad hoc microphone array can be useful for enhancing speech in such applications as teleconference systems without the need to prepare special devices. However, the positions of the sources (and the microphones when using an ad hoc microphone array) can change during recording, thus violating the commonly made assumption in many source separation algorithms that the mixing system is time-invariant. This paper proposes an extension of the multichannel nonnegative matrix factorization (NMF) approach to deal with the problem of underdetermined source separation in time-variant reverberant environments. The proposed method models the mixing system as a non-negative convolutive mixture based on the concept of a “semi-time-variant system” to handle the reverberation in a room as well allowing for relatively small changes in the source/microphone positions. It also models the power spectrogram of each sound source using the convolutive NMF model to consider the local dynamics of speech.},
  keywords = {deconvolution;matrix decomposition;microphone arrays;reverberation;source separation;speech enhancement;tensors;reverberation-robust underdetermined source separation;nonnegative tensor double deconvolution;ad hoc microphone array;speech enhancement;mixing system;multichannel nonnegative matrix factorization approach;time-variant reverberant environment;nonnegative convolutive mixture;semitime-variant system concept;power spectrogram;Spectrogram;Microphone arrays;Source separation;Time-frequency analysis;Speech;Europe},
  doi = {10.1109/EUSIPCO.2016.7760528},
  issn = {2076-1465},
  month = {Aug},
  url = {https://www.eurasip.org/proceedings/eusipco/eusipco2016/papers/1570252361.pdf},
}
Downloads: 0