LPSPECTRUM

LPSPECTRUM - LP smoothed spectrum

Usage:

LPSPECTRUM x n d typ c lfft

Inputs:

x	signal vector (nx samples)	no default
n	number of coefficients	def.=20
d	pre-emphasis (differentiation factor, 0..1)	def.=1
typ	selects the type of output spectrum	def.=LOGARITHMIC
c	spectrum scale factor or offset	def.=1(linear) or 0(dB)
lfft	fft length for spectrum computation	def.=2M>=nx

Outputs:

y	LPC-smoothed spectrum (vocal tract transfer function, lfft/2+1 samples)
alpha	error energy

Function:

This function computes the LPC-smoothed spectrum of the speech signal X. First the inverse filter coefficients are computed using the same method (differentiation and auto-correlation method) as in LPCOEFS and than the frequency response (amplitudes) for this filter is computed using an FFT. The signal vector x must be windowed before being connected to this atom. Optional pre-filtering (differentiation, if d>0) is applied to the signal before the coefficients' iteration. The spectrum computed using this method can be interpreted as the transfer function of the human vocal tract and is mainly used for speech formant (candidate) extraction.

Type of output spectrum selected by input typ

value of typ	type of spectrum stored in y	usage of c
0 or POWER	power spectrum	scaling factor
1 or LINEAR	amplitude spectrum	scaling factor
2 or LOGARITHMIC	logarithmic amplitude spectrum	offset in dB