Independent Low-Rank Tensor Analysis for Audio Source Separation. Yoshii, K., Kitamura, K., Bando, Y., Nakamura, E., & Kawahara, T. In 2018 26th European Signal Processing Conference (EUSIPCO), pages 1657-1661, Sep., 2018.
Independent Low-Rank Tensor Analysis for Audio Source Separation [pdf]Paper  doi  abstract   bibtex   
This paper describes a versatile tensor factorization technique called independent low-rank tensor analysis (ILRTA) and its application to single-channel audio source separation. In general, audio source separation has been conducted in the short-time Fourier transform (STFT) domain under an unrealistic but conventional assumption of the independence of time-frequency (TF) bins. Nonnegative matrix factorization (NMF) is a typical technique of single-channel source separation based on the low-rankness of source spectrograms. In a multichannel setting, independent component analysis (ICA) and its multivariate extension called independent vector analysis (IVA) have often been used for blind source separation based on the independence of source spectrograms. Integrating NMF and IVA, independent low-rank matrix analysis (ILRMA) was recently proposed. To deal with the covariance of TF bins, in this paper we propose ILRTA as a new extension of NMF. Both ILRMA and ILRTA aim to find independent and low-rank sources. A key difference is that while ILRMA estimates demixing filters that decorrelate the channels for multichannel source separation, ILRTA finds optimal transforms that decorrelate the time frames and frequency bins of a STFT representation for single-channel source separation in a way that the bin-wise independence assumed by NMF holds true as much as possible. We report evaluation results of ILRTA and discuss extension of ILRTA to multichannel source separation.
@InProceedings{8553013,
  author = {K. Yoshii and K. Kitamura and Y. Bando and E. Nakamura and T. Kawahara},
  booktitle = {2018 26th European Signal Processing Conference (EUSIPCO)},
  title = {Independent Low-Rank Tensor Analysis for Audio Source Separation},
  year = {2018},
  pages = {1657-1661},
  abstract = {This paper describes a versatile tensor factorization technique called independent low-rank tensor analysis (ILRTA) and its application to single-channel audio source separation. In general, audio source separation has been conducted in the short-time Fourier transform (STFT) domain under an unrealistic but conventional assumption of the independence of time-frequency (TF) bins. Nonnegative matrix factorization (NMF) is a typical technique of single-channel source separation based on the low-rankness of source spectrograms. In a multichannel setting, independent component analysis (ICA) and its multivariate extension called independent vector analysis (IVA) have often been used for blind source separation based on the independence of source spectrograms. Integrating NMF and IVA, independent low-rank matrix analysis (ILRMA) was recently proposed. To deal with the covariance of TF bins, in this paper we propose ILRTA as a new extension of NMF. Both ILRMA and ILRTA aim to find independent and low-rank sources. A key difference is that while ILRMA estimates demixing filters that decorrelate the channels for multichannel source separation, ILRTA finds optimal transforms that decorrelate the time frames and frequency bins of a STFT representation for single-channel source separation in a way that the bin-wise independence assumed by NMF holds true as much as possible. We report evaluation results of ILRTA and discuss extension of ILRTA to multichannel source separation.},
  keywords = {audio signal processing;blind source separation;Fourier transforms;independent component analysis;matrix decomposition;source separation;tensors;NMF;low-rank matrix analysis;ILRMA;ILRTA;low-rank sources;multichannel source separation;single-channel source separation;bin-wise independence;low-rank tensor analysis;single-channel audio source separation;time-frequency bins;nonnegative matrix factorization;source spectrograms;independent component analysis;independent vector analysis;blind source separation;tensor factorization technique;short-time Fourier transform;Covariance matrices;Spectrogram;Source separation;Tensile stress;Time-frequency analysis;Transforms;Decorrelation},
  doi = {10.23919/EUSIPCO.2018.8553013},
  issn = {2076-1465},
  month = {Sep.},
  url = {https://www.eurasip.org/proceedings/eusipco/eusipco2018/papers/1570436717.pdf},
}
Downloads: 0