Low-rank inverse-gamma @ IEEE ICASSP 2016 (Shangai)

Dionyssis Kounades-Bastian, Laurent Girin, Xavier Alameda-Pineda, Sharon Gannot and Radu Horaud

Our low-rank parametrization of inverse-gamma distributions for speech modeling has been accepted at IEEE ICASSP 2016. The main idea is to parametrize the power-spectral density with an NMF-like model in which, even if the parameters are low-rank, the PSD is not. Indeed, the entries of the PSD matrix follow an inverse-gamma distribution, and the parameters of this distribution concatenated in a matrix are low rank. However, since the PSD itself is a hidden distribution, it is not forced to be low-rank. This is a bit difficult to explain in two lines. I am sure Dionyssos will do a great job explaining this at ICASSP. Check the paper [1].

References

1. D. Kounades-Bastian, L. Girin, X. Alameda-Pineda, S. Gannot, and R. Horaud, “An inverse-gama source variance prior with factorized parametrization for audio source separation,” in IEEE International Conference on Audio, Speech and Signal Processing, Shangai, China, 2016, pp. 136-140. [ bib pdf ]
   

   @inproceedings{Kounades-ICASSP-2016,
     TITLE = {An inverse-gama source variance prior with factorized parametrization for audio source separation},
     AUTHOR = {Kounades-Bastian, Dionyssos and Girin, Laurent and Alameda-Pineda, Xavier and Gannot, Sharon and Horaud, Radu},
     BOOKTITLE = {IEEE International Conference on Audio, Speech and Signal Processing},
     YEAR = {2016},
     address = {Shangai, China},
     pages={136-140},
     pdf={http://xavirema.eu/wp-content/papercite-data/pdf/Kounades-ICASSP-2016.pdf}
   }