TSP (libtsp) - SPlpcISdist

SPlpcISdist

Routine

double SPlpcISdist (const float pc1[], const float pc2[], int Np)

Purpose

Calculate the Itakura-Saito LPC spectral distance measure

Description

This function calculates the Itakura-Saito LPC spectral distance measure kernel. Given two LPC spectral models 1/A1(z) and 1/A2(z), this function calculates a spectral difference function. Let the coefficient vectors corresponding to the all-pole filters be a1 and a2. Consider a signal with correlation corresponding to the correlation matrix R1. Define E1 as the minimum mean-square error that results when the signal is filtered using the optimum prediction filter A1(z) derived from the correlation R1. Define E12 as the mean-square error that results when the same signal is filtered using the prediction filter A2(z) (mismatched to the signal). These errors can be calculated as as

   E1  = a1' R1 a1 ,
   E12 = a2' R1 a2 ,

This function returns the value E12/E1.

The distance kernel can be used to calculate the Itakura-Saito maximum likelihood spectral distance between two spectra models g1/A1(z) and g2/A2(z),

          1    pi
  Dist = ---  Int [ exp(V(w)) - V(w) - 1 ] dw.
         2pi  -pi
                     g1^2             g2^2
  where  V(w) = ln ---------  - ln  --------- ,
                   |A1(w)|^2        |A2(w)|^2

   pi
  Int V(w) dw = 4 pi ln(g1/g2),
  -pi

   pi
  Int exp(V(w)) dw = (g1/g2)^2  E12/E1
  -pi

where E1 is the residual energy for A1(z) and E12 is the residual energy for A2(z). If the gains g1 and g2 are equal (ref. 1),

  Dist = E12/E1 -  1,

An alternate spectral difference measure is the Itakura measure (see ref. 2),

         a1' R2 a1
  d = ln --------- .
         a2' R2 a2

The term inside the logarithm is obtained by interchanging the roles of a1 and a2 in the call to this routine.

References:

A. Gray Jr. and J. Markel, "Distance measures for speech processing", IEEE Trans. Acoustics, Speech, and Signal Processing, vol. ASSP-24, pp. 380-391, October 1976.

M. R. Sambur and N.S. Jayant, "LPC analysis/synthesis from speech inputs containing quantizing noise or additive white noise", IEEE Trans. Acoustics, Speech, and Signal Processing, vol. ASSP-24, pp. 488-494, December 1976.

Parameters

<- double SPlpcISdist: Resultant distance measure
-> const float pc1[]: Predictor coefficients corresponding to the first spectrum
-> const float pc2[]: Predictor coefficients corresponding to the second spectrum
-> int Np: Number of predictor coefficients (maximum 50)

Author / revision