SPlpcLSdist

Routine

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

Purpose

Calculate the RMS log LPC spectral distance measure

Description

              Np       -k
  A(z) = 1 - SUM p(k) z  .
             k=1

The power spectrum corresponding to the all-pole LPC filter is

  S(w) = 1 / |A(w)|^2 ,

where A(w) is short-hand notation for A(exp(jw)). The RMS log LPC spectral distance measure for two filters A1(z) and A2(z) is given by

              pi
  D = sqrt{ int (10 log[|A1(w)|^2] - 10 log[|A2(w)|^2])^2 dw } .
             -pi

The power spectrum is the Fourier transform of the (infinite) set of autocorrelation coefficients. The cepstrum for the autocorrelation coefficients is given by inverse Fourier transform of the log power spectrum. Equivalently the cepstral coefficients are the Fourier series coefficients for the (periodic) log power spectrum.

The RMS log LPC spectral distance measure is equivalent to the root mean-square difference between the cepstral coefficients corresponding to the two LPC filters. This routine finds the cepstral coefficients corresponding to the two sets of input predictor coefficients. Theoretically, this requires an infinite number of cepstral coefficients. This routine calculates an approximation to the distance from two finite length sets of cepstral coefficients.

Reference:

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

Parameters

<- double SPlpcLSdist

Resultant distance measure (dB units). 0 dB corresponds to identical sets of predictor coefficients.

-> 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 in each of pc1 and pc2

-> int Ncep

Number of cepstral coefficients used for the distance calculation (maximum 150). This number does not include the zero'th cepstral coefficient. A recommended number is 3*Np.

Author / revision

See Also