preprocess_spx.c

一个开源SIP协议栈

/* Copyright (C) 2003 Epic Games (written by Jean-Marc Valin)
   Copyright (C) 2004-2006 Epic Games 
   
   File: preprocess.c
   Preprocessor with denoising based on the algorithm by Ephraim and Malah

   Redistribution and use in source and binary forms, with or without
   modification, are permitted provided that the following conditions are
   met:

   1. Redistributions of source code must retain the above copyright notice,
   this list of conditions and the following disclaimer.

   2. Redistributions in binary form must reproduce the above copyright
   notice, this list of conditions and the following disclaimer in the
   documentation and/or other materials provided with the distribution.

   3. The name of the author may not be used to endorse or promote products
   derived from this software without specific prior written permission.

   THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
   IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
   OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
   DISCLAIMED. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT,
   INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
   (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
   SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
   HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
   STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN
   ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
   POSSIBILITY OF SUCH DAMAGE.
*/


/*
   Recommended papers:
   
   Y. Ephraim and D. Malah, "Speech enhancement using minimum mean-square error
   short-time spectral amplitude estimator". IEEE Transactions on Acoustics, 
   Speech and Signal Processing, vol. ASSP-32, no. 6, pp. 1109-1121, 1984.
   
   Y. Ephraim and D. Malah, "Speech enhancement using minimum mean-square error
   log-spectral amplitude estimator". IEEE Transactions on Acoustics, Speech and 
   Signal Processing, vol. ASSP-33, no. 2, pp. 443-445, 1985.
   
   I. Cohen and B. Berdugo, "Speech enhancement for non-stationary noise environments".
   Signal Processing, vol. 81, no. 2, pp. 2403-2418, 2001.

   Stefan Gustafsson, Rainer Martin, Peter Jax, and Peter Vary. "A psychoacoustic 
   approach to combined acoustic echo cancellation and noise reduction". IEEE 
   Transactions on Speech and Audio Processing, 2002.
   
   J.-M. Valin, J. Rouat, and F. Michaud, "Microphone array post-filter for separation
   of simultaneous non-stationary sources". In Proceedings IEEE International 
   Conference on Acoustics, Speech, and Signal Processing, 2004.
*/

#ifdef HAVE_CONFIG_H
#include "config.h"
#endif

#include <math.h>
#include "speex/speex_preprocess.h"
#include "speex/speex_echo.h"
#include "misc.h"
#include "fftwrap.h"
#include "filterbank.h"
#include "math_approx.h"

#ifndef M_PI
#define M_PI 3.14159263
#endif

#define LOUDNESS_EXP 2.5

#define NB_BANDS 24

#define SPEECH_PROB_START_DEFAULT       QCONST16(0.35f,15)
#define SPEECH_PROB_CONTINUE_DEFAULT    QCONST16(0.20f,15)
#define NOISE_SUPPRESS_DEFAULT       -25
#define ECHO_SUPPRESS_DEFAULT        -45
#define ECHO_SUPPRESS_ACTIVE_DEFAULT -15

#ifndef NULL
#define NULL 0
#endif

#define SQR(x) ((x)*(x))
#define SQR16(x) (MULT16_16((x),(x)))
#define SQR16_Q15(x) (MULT16_16_Q15((x),(x)))

#ifdef FIXED_POINT
static inline spx_word16_t DIV32_16_Q8(spx_word32_t a, spx_word32_t b)
{
   if (SHR32(a,7) >= b)
   {
      return 32767;
   } else {
      if (b>=QCONST32(1,23))
      {
         a = SHR32(a,8);
         b = SHR32(b,8);
      }
      if (b>=QCONST32(1,19))
      {
         a = SHR32(a,4);
         b = SHR32(b,4);
      }
      if (b>=QCONST32(1,15))
      {
         a = SHR32(a,4);
         b = SHR32(b,4);
      }
      a = SHL32(a,8);
      return PDIV32_16(a,b);
   }
   
}
static inline spx_word16_t DIV32_16_Q15(spx_word32_t a, spx_word32_t b)
{
   if (SHR32(a,15) >= b)
   {
      return 32767;
   } else {
      if (b>=QCONST32(1,23))
      {
         a = SHR32(a,8);
         b = SHR32(b,8);
      }
      if (b>=QCONST32(1,19))
      {
         a = SHR32(a,4);
         b = SHR32(b,4);
      }
      if (b>=QCONST32(1,15))
      {
         a = SHR32(a,4);
         b = SHR32(b,4);
      }
      a = SHL32(a,15)-a;
      return DIV32_16(a,b);
   }
}
#define SNR_SCALING 256.f
#define SNR_SCALING_1 0.0039062f
#define SNR_SHIFT 8

#define FRAC_SCALING 32767.f
#define FRAC_SCALING_1 3.0518e-05
#define FRAC_SHIFT 1

#define EXPIN_SCALING 2048.f
#define EXPIN_SCALING_1 0.00048828f
#define EXPIN_SHIFT 11
#define EXPOUT_SCALING_1 1.5259e-05

#define NOISE_SHIFT 7

#else

#define DIV32_16_Q8(a,b) ((a)/(b))
#define DIV32_16_Q15(a,b) ((a)/(b))
#define SNR_SCALING 1.f
#define SNR_SCALING_1 1.f
#define SNR_SHIFT 0
#define FRAC_SCALING 1.f
#define FRAC_SCALING_1 1.f
#define FRAC_SHIFT 0
#define NOISE_SHIFT 0

#define EXPIN_SCALING 1.f
#define EXPIN_SCALING_1 1.f
#define EXPOUT_SCALING_1 1.f

#endif

/** Speex pre-processor state. */
struct SpeexPreprocessState_ {
   /* Basic info */
   int    frame_size;        /**< Number of samples processed each time */
   int    ps_size;           /**< Number of points in the power spectrum */
   int    sampling_rate;     /**< Sampling rate of the input/output */
   int    nbands;
   FilterBank *bank;
   
   /* Parameters */
   int    denoise_enabled;
   int    agc_enabled;
   float  agc_level;
   int    vad_enabled;
   int    dereverb_enabled;
   spx_word16_t  reverb_decay;
   spx_word16_t  reverb_level;
   spx_word16_t speech_prob_start;
   spx_word16_t speech_prob_continue;
   int    noise_suppress;
   int    echo_suppress;
   int    echo_suppress_active;
   SpeexEchoState *echo_state;
   
   /* DSP-related arrays */
   spx_word16_t *frame;      /**< Processing frame (2*ps_size) */
   spx_word16_t *ft;         /**< Processing frame in freq domain (2*ps_size) */
   spx_word32_t *ps;         /**< Current power spectrum */
   spx_word16_t *gain2;      /**< Adjusted gains */
   spx_word16_t *gain_floor; /**< Minimum gain allowed */
   spx_word16_t *window;     /**< Analysis/Synthesis window */
   spx_word32_t *noise;      /**< Noise estimate */
   spx_word32_t *reverb_estimate; /**< Estimate of reverb energy */
   spx_word32_t *old_ps;     /**< Power spectrum for last frame */
   spx_word16_t *gain;       /**< Ephraim Malah gain */
   spx_word16_t *prior;      /**< A-priori SNR */
   spx_word16_t *post;       /**< A-posteriori SNR */

   spx_word32_t *S;          /**< Smoothed power spectrum */
   spx_word32_t *Smin;       /**< See Cohen paper */
   spx_word32_t *Stmp;       /**< See Cohen paper */
   int *update_prob;       /**< Propability of speech presence for noise update */

   spx_word16_t *zeta;       /**< Smoothed a priori SNR */
   spx_word32_t *echo_noise;
   spx_word32_t *residual_echo;

   /* Misc */
   spx_word16_t *inbuf;      /**< Input buffer (overlapped analysis) */
   spx_word16_t *outbuf;     /**< Output buffer (for overlap and add) */

#ifndef FIXED_POINT
   float *loudness_weight;   /**< Perceptual loudness curve */
   float  loudness;          /**< loudness estimate */
   float  loudness2;         /**< loudness estimate */
   int    nb_loudness_adapt; /**< Number of frames used for loudness adaptation so far */
#endif
   int    nb_adapt;          /**< Number of frames used for adaptation so far */
   int    was_speech;
   int    min_count;         /**< Number of frames processed so far */
   void  *fft_lookup;        /**< Lookup table for the FFT */
#ifdef FIXED_POINT
   int    frame_shift;
#endif
};


static void conj_window(spx_word16_t *w, int len)
{
   int i;
   for (i=0;i<len;i++)
   {
      spx_word16_t tmp;
      spx_word16_t x = DIV32_16(MULT16_16(QCONST16(4.f,13),i),len);
      int inv=0;
      if (x<QCONST16(1.f,13))
      {
      } else if (x<QCONST16(2.f,13))
      {
         x=QCONST16(2.f,13)-x;
         inv=1;
      } else if (x<QCONST16(3.f,13))
      {
         x=x-QCONST16(2.f,13);
         inv=1;
      } else {
         x=QCONST16(2.f,13)-x+QCONST16(2.f,13); /* 4 - x */
      }
      x = MULT16_16_Q14(QCONST16(1.271903f,14), x);
      tmp = SQR16_Q15(QCONST16(.5f,15)-MULT16_16_P15(QCONST16(.5f,15),spx_cos_norm(QCONST32(x,2))));
      if (inv)
         tmp=SUB16(Q15_ONE,tmp);
      w[i]=spx_sqrt(SHL32(EXTEND32(tmp),15));
   }
}

      
#ifdef FIXED_POINT
/* This function approximates the gain function 
   y = gamma(1.25)^2 * M(-.25;1;-x) / sqrt(x)  
   which multiplied by xi/(1+xi) is the optimal gain
   in the loudness domain ( sqrt[amplitude] )
   Input in Q11 format, output in Q15
*/
static inline spx_word32_t hypergeom_gain(spx_word32_t xx)
{
   int ind;
   spx_word16_t frac;
   /* Q13 table */
   static const spx_word16_t table[21] = {
       6730,  8357,  9868, 11267, 12563, 13770, 14898,
      15959, 16961, 17911, 18816, 19682, 20512, 21311,
      22082, 22827, 23549, 24250, 24931, 25594, 26241};
      ind = SHR32(xx,10);
      if (ind<0)
         return Q15_ONE;
      if (ind>19)
         return ADD32(EXTEND32(Q15_ONE),EXTEND32(DIV32_16(QCONST32(.1296,23), SHR32(xx,EXPIN_SHIFT-SNR_SHIFT))));
      frac = SHL32(xx-SHL32(ind,10),5);
      return SHL32(DIV32_16(PSHR32(MULT16_16(Q15_ONE-frac,table[ind]) + MULT16_16(frac,table[ind+1]),7),(spx_sqrt(SHL32(xx,15)+6711))),7);
}

static inline spx_word16_t qcurve(spx_word16_t x)
{
   x = MAX16(x, 1);
   return DIV32_16(SHL32(EXTEND32(32767),9),ADD16(512,MULT16_16_Q15(QCONST16(.60f,15),DIV32_16(32767,x))));
}

/* Compute the gain floor based on different floors for the background noise and residual echo */
static void compute_gain_floor(int noise_suppress, int effective_echo_suppress, spx_word32_t *noise, spx_word32_t *echo, spx_word16_t *gain_floor, int len)
{
   int i;
   
   if (noise_suppress > effective_echo_suppress)
   {
      spx_word16_t noise_gain, gain_ratio;
      noise_gain = EXTRACT16(MIN32(Q15_ONE,SHR32(spx_exp(MULT16_16(QCONST16(0.11513,11),noise_suppress)),1)));
      gain_ratio = EXTRACT16(MIN32(Q15_ONE,SHR32(spx_exp(MULT16_16(QCONST16(.2302585f,11),effective_echo_suppress-noise_suppress)),1)));

      /* gain_floor = sqrt [ (noise*noise_floor + echo*echo_floor) / (noise+echo) ] */
      for (i=0;i<len;i++)
         gain_floor[i] = MULT16_16_Q15(noise_gain,
                                       spx_sqrt(SHL32(EXTEND32(DIV32_16_Q15(PSHR32(noise[i],NOISE_SHIFT) + MULT16_32_Q15(gain_ratio,echo[i]),
                                             (1+PSHR32(noise[i],NOISE_SHIFT) + echo[i]) )),15)));
   } else {
      spx_word16_t echo_gain, gain_ratio;
      echo_gain = EXTRACT16(MIN32(Q15_ONE,SHR32(spx_exp(MULT16_16(QCONST16(0.11513,11),effective_echo_suppress)),1)));
      gain_ratio = EXTRACT16(MIN32(Q15_ONE,SHR32(spx_exp(MULT16_16(QCONST16(.2302585f,11),noise_suppress-effective_echo_suppress)),1)));

      /* gain_floor = sqrt [ (noise*noise_floor + echo*echo_floor) / (noise+echo) ] */
      for (i=0;i<len;i++)
         gain_floor[i] = MULT16_16_Q15(echo_gain,
                                       spx_sqrt(SHL32(EXTEND32(DIV32_16_Q15(MULT16_32_Q15(gain_ratio,PSHR32(noise[i],NOISE_SHIFT)) + echo[i],
                                             (1+PSHR32(noise[i],NOISE_SHIFT) + echo[i]) )),15)));
   }
}

#else
/* This function approximates the gain function 
   y = gamma(1.25)^2 * M(-.25;1;-x) / sqrt(x)  
   which multiplied by xi/(1+xi) is the optimal gain
   in the loudness domain ( sqrt[amplitude] )
*/
static inline spx_word32_t hypergeom_gain(spx_word32_t xx)
{
   int ind;
   float integer, frac;
   float x;
   static const float table[21] = {
      0.82157f, 1.02017f, 1.20461f, 1.37534f, 1.53363f, 1.68092f, 1.81865f,
      1.94811f, 2.07038f, 2.18638f, 2.29688f, 2.40255f, 2.50391f, 2.60144f,
      2.69551f, 2.78647f, 2.87458f, 2.96015f, 3.04333f, 3.12431f, 3.20326f};
      x = EXPIN_SCALING_1*xx;
      integer = floor(2*x);
      ind = (int)integer;
      if (ind<0)
         return FRAC_SCALING;
      if (ind>19)
         return FRAC_SCALING*(1+.1296/x);
      frac = 2*x-integer;
      return FRAC_SCALING*((1-frac)*table[ind] + frac*table[ind+1])/sqrt(x+.0001f);
}

static inline spx_word16_t qcurve(spx_word16_t x)
{
   return 1.f/(1.f+.15f/(SNR_SCALING_1*x));
}

static void compute_gain_floor(int noise_suppress, int effective_echo_suppress, spx_word32_t *noise, spx_word32_t *echo, spx_word16_t *gain_floor, int len)
{
   int i;
   float echo_floor;
   float noise_floor;

   noise_floor = exp(.2302585f*noise_suppress);
   echo_floor = exp(.2302585f*effective_echo_suppress);

   /* Compute the gain floor based on different floors for the background noise and residual echo */
   for (i=0;i<len;i++)
      gain_floor[i] = FRAC_SCALING*sqrt(noise_floor*PSHR32(noise[i],NOISE_SHIFT) + echo_floor*echo[i])/sqrt(1+PSHR32(noise[i],NOISE_SHIFT) + echo[i]);
}

#endif
SpeexPreprocessState *speex_preprocess_state_init(int frame_size, int sampling_rate)
{
   int i;