To find the Linear predictive coefficents all are using auto correlation with levinson durbin equation. I want to know is there any alternative method availble that still reduces the computational complexity to find the LPC coefficients.

# Linear predictive coefficient computation

Started by ●November 29, 2005

Reply by ●November 29, 20052005-11-29

RAMU wrote:> To find the Linear predictive coefficents all are using auto > correlation with levinson durbin equation. I want to know is there any > alternative method availble that still reduces the computational > complexity to find the LPC coefficients.Hello Ramu, I guess this is stadardized to use Levinson-Durbin for LPC solution. There may be Schur algorithm as follows that might give faster computation. --- Schur's recursion from 1917 is related to the Levinson-Durbin method, * but faster on parallel architectures; where Levinson-Durbin would take time * proportional to p * log(p), Schur only requires time proportional to p. The * GSM coder uses an integer version of the Schur recursion. If you search in net you may find Schur recusion algorithm. Rgs, Santosh

Reply by ●November 30, 20052005-11-30

RAMU wrote:> To find the Linear predictive coefficents all are using auto > correlation with levinson durbin equation. I want to know is there any > alternative method availble that still reduces the computational > complexity to find the LPC coefficients.If you are willing to trade filter optimality with computation speed, you can use an adaptive method and stop after a few iterations. You can tune the adaption parameter for fast tracking. GAL LMS has the advantage that the convergence rate is roughly independent of the signal statistics (eigenspread). I also remember reading about an adaptive method that used a cascade of FIRs (instead of one big filter). The pro was extremely fast convergence and low computational complexity. As usual, you pay for that with accuracy. They suggested the use of cascade FIRs for fast and good initial estimates for standard adaptive methods. Can't remember the exact title of the paper. As a last resort, you might want to look into an adaptive IIR. Regards, Andor