From: HardySpicer on 11 Sep 2009 05:30 Ok so I have been trying to separate a convolutive mixture of two speech signals using the standard literature. (Blind source separation  Natural Gradient algorithm). I find it works ok but have reservations. Often I can estimate one of the channels and not the second and at other times I can estimate the second and not the first. Sometimes if I am lucky I can get both! Also there does not appear to be a normalized version of the algorithm like LMS so getting stability is a wild guess for the step size. It also appears to go through local nulls...ie it gets an estimate then it gets worse then better again etc. Anybody have experience in this? Hardy
From: maury001 on 11 Sep 2009 11:09 On Sep 11, 4:30 am, HardySpicer <gyansor...(a)gmail.com> wrote: > Ok so I have been trying to separate a convolutive mixture of two > speech signals using the standard literature. (Blind source separation >  Natural Gradient algorithm). I find it works ok but have > reservations. Often I can estimate one of the channels and not the > second and at other times I can estimate the second and not the first. > Sometimes if I am lucky I can get both! Also there does not appear to > be a normalized version of the algorithm like LMS so getting stability > is a wild guess for the step size. It also appears to go through local > nulls...ie it gets an estimate then it gets worse then better again > etc. Anybody have experience in this? > > Hardy Look ay a paper by Aapo Hyvärinen and Erkki Oja: "Independent Component Analysis: Algorithms and Applications", and contrast that with one by TeWon Lee : "Independent Component Analysis Using an Extended Infomax Algorithm for Mixed Subgaussian and Supergaussian Sources". Hyvärinen uses matrix transformation to rotate the axis of the signal to cause the mixed signals to be independent after rotation. Lee uses a learning algorithm, and discusses the behavior. Maurice Givens
From: HardySpicer on 11 Sep 2009 15:33 On Sep 11, 8:09 am, maury...(a)core.com wrote: > On Sep 11, 4:30 am, HardySpicer <gyansor...(a)gmail.com> wrote: > > > Ok so I have been trying to separate a convolutive mixture of two > > speech signals using the standard literature. (Blind source separation > >  Natural Gradient algorithm). I find it works ok but have > > reservations. Often I can estimate one of the channels and not the > > second and at other times I can estimate the second and not the first. > > Sometimes if I am lucky I can get both! Also there does not appear to > > be a normalized version of the algorithm like LMS so getting stability > > is a wild guess for the step size. It also appears to go through local > > nulls...ie it gets an estimate then it gets worse then better again > > etc. Anybody have experience in this? > > > Hardy > > Look ay a paper by Aapo Hyvärinen and Erkki Oja: "Independent > Component Analysis: Algorithms and Applications", and contrast that > with one by TeWon Lee : "Independent Component Analysis Using an > Extended Infomax Algorithm for Mixed Subgaussian and Supergaussian > Sources". Hyvärinen uses matrix transformation to rotate the axis of > the signal to cause the mixed signals to be independent after > rotation. Lee uses a learning algorithm, and discusses the behavior. > > Maurice Givens Thanks for that. Unfortunately they are using a fixed matrix as the mixer. In real life we need a polynomial matrix to account for reverberation and the like. Of course it can be extended. Hardy

Pages: 1 Prev: Top 3 DSP/SP books Next: FDL / DASIP / S4D 2009 > Invitation to participate 