postfilt.m

FS1016源代码

% MATLAB SIMULATION OF NSA FS-1016 CELP v3.2
% COPYRIGHT (C) 1995-99 ANDREAS SPANIAS AND TED PAINTER
%
% This Copyright applies only to this particular MATLAB implementation
% of the FS-1016 CELP coder.  The MATLAB software is intended only for educational
% purposes.  No other use is intended or authorized.  This is not a public
% domain program and distribution to individuals or networks is strictly
% prohibited.  Be aware that use of the standard in any form is goverened
% by rules of the US DoD.  Therefore patents and royalties may apply to
% authors, companies, or committees associated with this standard, FS-1016.  For
% questions regarding the MATLAB implementation please contact Andreas
% Spanias at [phone].  For questions on rules,
% royalties, or patents associated with the standard, please contact the DoD.
%
% ALL DERIVATIVE WORKS MUST INCLUDE THIS COPYRIGHT NOTICE.
%
% ******************************************************************
% POSTFILT
%
% PORTED TO MATLAB FROM CELP 3.2a C RELEASE
% 8-9-94
%
% ******************************************************************
%
% DESCRIPTION
%
% Reduce coder noise by applying an adaptive postfilter to output speech
%
% DESIGN NOTES
%
% Adaptive postfilter routine to reduce perceptual coder noise.
% This filter emphasizes spectral regions predicted by
% short-term LPC analysis, which tends to mask coder noise by
% concentrating it under the formant peaks.  Unfortunately, acoustic
% background noise may also be enhanced because LPC analysis often
% models acoustic noise instead of speech formants.  In addtion,
% postfiltering can be detrimental to tandem coding if not taken
% into consideration.  To overcome these problems, FS1016 SHOULD
% eventually incorporate the postfilter's enhancement properties
% into the analysis process.
%
% Adaptive spectral tilt compensation is applied to flatten the
% overall tilt of the postfilter.  Slight high frequency boost is
% applied for output shaping.  A pitch postfilter is used to reduce
% pitch buzz.  Finally, AGC compensates for the filter gains using
% a time constant set by parameter TC that should be dependent on
% frame length.
%
% REFERENCES
%
% 1. Chen & Gersho, Real-Time Vector APC Speech Coding at 4800 bps
%    with Adaptive Postfiltering, ICASSP '87
%
% 2. Juin-Hwey (Raymond) Chen, "Low-Bit-Rate Predictive Coding of
%    Speech Waveforms Based on Vector Quantization," PhD Dissertation,
%    UCSB ECE Dept., March 1987.
%
% 3. Ramamoorthy, Jayant, Cox, & Sondhi, "Enhancement of ADPCM Speech
%    Coding with Backward-Adaptive Algorithms for Postfiltering and
%    Noise Feedback," IEEE JOSAIC, Feb. 1988, pp. 364-382.
%
% VARIABLES
%
% INPUTS
%   s           -     Input speech (coder output,non-postfiltered)
%   l           -     Subframe size
%   alpha       -     Filter parameter
%   beta        -     Filter parameter
%   powerin     -     Input power estimate
%   powerout    -     Output power estimate
%   dp1         -     Filter memory
%   dp2         -     Filter memory
%   dp3         -     Filter memory
%   fci         -     LPC predictor poles
%   no          -     Predictor order
%
% OUTPUTS
%   s           -     Output speech (postfiltered)
%   powerin     -     Input power estimate
%   powerout    -     Output power estimate
%   dp1         -     Filter memory
%   dp2         -     Filter memory
%   dp3         -     Filter memory
%
% INTERNALS
%   pcexp1      -     Bandwidth expanded predictor poles = postfilter zeros
%   pcexp2      -     Bandwidth expanded predictor poles = postfilter poles
%   rcexp2      -     Reflection coefficients derived from postfilter poles
%   newpowerin  -     Input power estimate vector
%   newpowerout -     Output power estimate vector
%   agcScale    -     Automatic gain control scaling vector
%   agcUnity    -     Unity gain locations in agc vector
%   ast         -     Spectral tilt compensation filter memory
%
% GLOBALS
%   ipZ         -     Input power estimation filter memory
%   opZ         -     Output power estimation filter memory
%
% CONSTANTS
%   TC          -     Input/Output power estimate time constant
%
% ******************************************************************

function [ s, powerin, powerout, dp1, dp2, dp3 ] = ...
         postfilt( s, l, alpha, beta, powerin, powerout, dp1, dp2, ...
                   dp3, fci, no )

% DECLARE GLOBAL CONSTANTS
global TC

% DECLARE GLOBAL VARIABLES
global ipZ opZ

% ALLOCATE LOCAL FILTER MEMORY (SPECTRAL TILT COMPENSATOR)
ast = zeros( 2, 1 );

% ESTIMATE INPUT POWER
[ newpowerin, ipZ ] = filter( [TC 0], [1 -1+TC], ( s .* s ), ipZ );
powerin = newpowerin(l);

% DO BANDWIDTH EXPANSION ON PREDICTOR POLES TO FORM PERCEPTUAL POSTFILTER
pcexp1 = bwexp( beta, fci, no );
pcexp2 = bwexp( alpha, fci, no );

% APPLY POLE-ZERO POSTFILTER
[ dp1, s ] = zerofilt( pcexp1, no, dp1, s, l );
[ dp2, s ] = polefilt( pcexp2, no, dp2, s, l );

% ESTIMATE SPECTRAL TILT (1ST ORDER FIT) OF POSTFILTER
% DENOMINATOR DOMINATES (POLES)
rcexp2 = pctorc( pcexp2, no );

% ADD TILT COMPENSATION BY A SCALED ZERO (DON'T ALLOW HF ROLL-OFF)
ast(1) = 1.0;
if rcexp2(1) > 0.0
    ast(2) = -0.5 * rcexp2(1);
else
    ast(2) = 0.00;
end
[ dp3, s ] = zerofilt( ast, 1, dp3, s, l );

% ESTIMATE OUTPUT POWER
[ newpowerout, opZ ] = filter( [TC 0], [1 -1+TC], ( s .* s ), opZ );
powerout = newpowerout(l);

% INCORPORATE SAMPLE-BY-SAMPLE AUTOMATIC GAIN CONTROL
agcUnity = find( newpowerout <= 0 );
newpowerout( agcUnity ) = ones( length( agcUnity ), 1 );
agcScale = sqrt( newpowerin ./ newpowerout );
agcScale( agcUnity ) = ones( length( agcUnity ), 1 );
s = s .* agcScale;