Hi, I am not sure if it is appropriate to post this here, if it isn't, can someone point me to a suitable forum on the net that can help. I have been reading the Voice Puppetry paper by Matthew Brand: http://www.merl.com/reports/docs/TR99-20.pdf but I am having some trouble understanding section 4.5 and especially equation 6. It is suppose to solve for a set of points Y but it seems to me to only solve for a single point y_t. Does that mean it needs to be applied iteratively to obtain the rest of the points or does it only represent part of the solution? any help will be appreciated. thanks

Quote:Original post by shaobohou
Hi,

I am not sure if it is appropriate to post this here, if it isn't, can someone point me to a suitable forum on the net that can help.

I have been reading the Voice Puppetry paper by Matthew Brand: http://www.merl.com/reports/docs/TR99-20.pdf
but I am having some trouble understanding section 4.5 and especially equation 6. It is suppose to solve for a set of points Y but it seems to me to only solve for a single point y_t. Does that mean it needs to be applied iteratively to obtain the rest of the points or does it only represent part of the solution?

any help will be appreciated.

thanks

I didn't have time to read the whole thing, but you appear to be correct in that it only solves for a single point y(t). However, the paper implies that at this point you already know s(t), the state space, from which you can compute any of the other y(t). Again, I didn't read it very thoroughly.

Quote:Original post by kSquared
I didn't have time to read the whole thing, but you appear to be correct in that it only solves for a single point y(t). However, the paper implies that at this point you already know s(t), the state space, from which you can compute any of the other y(t). Again, I didn't read it very thoroughly.

thanks for reply, I think there are lot of details not explained in the paper, at least they are not obvious to me. The blocked-banded system of linear equations derived is odd (I can see the blocks but where are the bands), and it seems to try to solve for the mean of the gaussians even though they are already known.

Equation 6, after all the transposes have been applied, is a D * 5D matrix multiplying a length 5D column vector and equals a length 5D column vector of zeroes, D = dim(y_t), I think. I am not sure how this non square system is solved, the section of the book it references only deal with LU decomposition of squared banded matrix.

this is frustrating.