ltp.c

TCPMP 0.72rc2已经添加FLV的支持

      for (i=0;i<3;i++)
         corr[i]=inner_prod(y[i],t,nsf);
      
      for (i=0;i<3;i++)
         for (j=0;j<=i;j++)
            A[i][j]=A[j][i]=inner_prod(y[i],y[j],nsf);
   }
#else
   {
      for (i=0;i<3;i++)
         corr[i]=inner_prod(x[i],target,nsf);
      
      for (i=0;i<3;i++)
         for (j=0;j<=i;j++)
            A[i][j]=A[j][i]=inner_prod(x[i],x[j],nsf);
   }
#endif

   {
      spx_word32_t C[9];
      const signed char *ptr=gain_cdbk;
      int best_cdbk=0;
      spx_word32_t best_sum=0;
      C[0]=corr[2];
      C[1]=corr[1];
      C[2]=corr[0];
      C[3]=A[1][2];
      C[4]=A[0][1];
      C[5]=A[0][2];      
      C[6]=A[2][2];
      C[7]=A[1][1];
      C[8]=A[0][0];
      
      /*plc_tuning *= 2;*/
      if (plc_tuning<2)
         plc_tuning=2;
#ifdef FIXED_POINT
      C[0] = MAC16_32_Q15(C[0],MULT16_16_16(plc_tuning,-327),C[0]);
      C[1] = MAC16_32_Q15(C[1],MULT16_16_16(plc_tuning,-327),C[1]);
      C[2] = MAC16_32_Q15(C[2],MULT16_16_16(plc_tuning,-327),C[2]);
      C[0] = SHL32(C[0],1);
      C[1] = SHL32(C[1],1);
      C[2] = SHL32(C[2],1);
      C[3] = SHL32(C[3],1);
      C[4] = SHL32(C[4],1);
      C[5] = SHL32(C[5],1);
#else
      C[0]*=1-.01*plc_tuning;
      C[1]*=1-.01*plc_tuning;
      C[2]*=1-.01*plc_tuning;
      C[6]*=.5*(1+.01*plc_tuning);
      C[7]*=.5*(1+.01*plc_tuning);
      C[8]*=.5*(1+.01*plc_tuning);
#endif
      for (i=0;i<gain_cdbk_size;i++)
      {
         spx_word32_t sum=0;
         spx_word16_t g[3];
         spx_word16_t pitch_control=64;
         spx_word16_t gain_sum;
         
         ptr = gain_cdbk+3*i;
         g[0]=ADD16((spx_word16_t)ptr[0],32);
         g[1]=ADD16((spx_word16_t)ptr[1],32);
         g[2]=ADD16((spx_word16_t)ptr[2],32);

         /* We favor "safe" pitch values to handle packet loss better */
         gain_sum = ADD16(ADD16(g[1],MAX16(g[0], 0)),MAX16(g[2], 0));
         if (gain_sum > 64)
         {
            gain_sum = SUB16(gain_sum, 64);
            if (gain_sum > 127)
               gain_sum = 127;
#ifdef FIXED_POINT
            pitch_control =  SUB16(64,EXTRACT16(PSHR32(MULT16_16(64,MULT16_16_16(plc_tuning, gain_sum)),10)));
#else
            pitch_control = 64*(1.-.001*plc_tuning*gain_sum);
#endif
            if (pitch_control < 0)
               pitch_control = 0;
         }
         
         sum = compute_pitch_error(C, g, pitch_control);
         
         if (sum>best_sum || i==0)
         {
            best_sum=sum;
            best_cdbk=i;
         }
      }
#ifdef FIXED_POINT
      gain[0] = ADD16(32,(spx_word16_t)gain_cdbk[best_cdbk*3]);
      gain[1] = ADD16(32,(spx_word16_t)gain_cdbk[best_cdbk*3+1]);
      gain[2] = ADD16(32,(spx_word16_t)gain_cdbk[best_cdbk*3+2]);
      /*printf ("%d %d %d %d\n",gain[0],gain[1],gain[2], best_cdbk);*/
#else
      gain[0] = 0.015625*gain_cdbk[best_cdbk*3]  + .5;
      gain[1] = 0.015625*gain_cdbk[best_cdbk*3+1]+ .5;
      gain[2] = 0.015625*gain_cdbk[best_cdbk*3+2]+ .5;
#endif
      *cdbk_index=best_cdbk;
   }

#ifdef FIXED_POINT
   for (i=0;i<nsf;i++)
     exc[i]=SHL32(ADD32(ADD32(MULT16_32_Q15(SHL16(gain[0],7),e[2][i]), MULT16_32_Q15(SHL16(gain[1],7),e[1][i])),
                        MULT16_32_Q15(SHL16(gain[2],7),e[0][i])), 2);
   
   err=0;
   for (i=0;i<nsf;i++)
   {
      spx_word16_t perr2;
      spx_sig_t tmp = SHL32(ADD32(ADD32(MULT16_32_Q15(SHL16(gain[0],7),x[2][i]),MULT16_32_Q15(SHL16(gain[1],7),x[1][i])),
                                  MULT16_32_Q15(SHL16(gain[2],7),x[0][i])),2);
      spx_sig_t perr=SUB32(target[i],tmp);
      new_target[i] = SUB32(target[i], tmp);
      perr2 = EXTRACT16(PSHR32(perr,15));
      err = ADD64(err,MULT16_16(perr2,perr2));
      
   }
#else
   for (i=0;i<nsf;i++)
      exc[i]=gain[0]*e[2][i]+gain[1]*e[1][i]+gain[2]*e[0][i];
   
   err=0;
   for (i=0;i<nsf;i++)
   {
      spx_sig_t tmp = gain[2]*x[0][i]+gain[1]*x[1][i]+gain[0]*x[2][i];
      new_target[i] = target[i] - tmp;
      err+=new_target[i]*new_target[i];
   }
#endif

   return err;
}


/** Finds the best quantized 3-tap pitch predictor by analysis by synthesis */
int pitch_search_3tap(
spx_sig_t target[],                 /* Target vector */
spx_sig_t *sw,
spx_coef_t ak[],                     /* LPCs for this subframe */
spx_coef_t awk1[],                   /* Weighted LPCs #1 for this subframe */
spx_coef_t awk2[],                   /* Weighted LPCs #2 for this subframe */
spx_sig_t exc[],                    /* Excitation */
const void *par,
int   start,                    /* Smallest pitch value allowed */
int   end,                      /* Largest pitch value allowed */
spx_word16_t pitch_coef,               /* Voicing (pitch) coefficient */
int   p,                        /* Number of LPC coeffs */
int   nsf,                      /* Number of samples in subframe */
SpeexBits *bits,
char *stack,
spx_sig_t *exc2,
spx_word16_t *r,
int complexity,
int cdbk_offset,
int plc_tuning
)
{
   int i,j;
   int cdbk_index, pitch=0, best_gain_index=0;
   VARDECL(spx_sig_t *best_exc);
   VARDECL(spx_sig_t *new_target);
   VARDECL(spx_sig_t *best_target);
   int best_pitch=0;
   spx_word64_t err, best_err=-1;
   int N;
   const ltp_params *params;
   VARDECL(int *nbest);

   N=complexity;
   if (N>10)
      N=10;
   if (N<1)
      N=1;

   ALLOC(nbest, N, int);
   params = (const ltp_params*) par;

   if (end<start)
   {
      speex_bits_pack(bits, 0, params->pitch_bits);
      speex_bits_pack(bits, 0, params->gain_bits);
      for (i=0;i<nsf;i++)
         exc[i]=0;
      return start;
   }
   
   ALLOC(best_exc, nsf, spx_sig_t);
   ALLOC(new_target, nsf, spx_sig_t);
   ALLOC(best_target, nsf, spx_sig_t);
   
   if (N>end-start+1)
      N=end-start+1;
   if (end != start)
      open_loop_nbest_pitch(sw, start, end, nsf, nbest, NULL, N, stack);
   else
      nbest[0] = start;
   for (i=0;i<N;i++)
   {
      pitch=nbest[i];
      for (j=0;j<nsf;j++)
         exc[j]=0;
      err=pitch_gain_search_3tap(target, ak, awk1, awk2, exc, par, pitch, p, nsf,
                                 bits, stack, exc2, r, new_target, &cdbk_index, cdbk_offset, plc_tuning);
      if (err<best_err || best_err<0)
      {
         for (j=0;j<nsf;j++)
            best_exc[j]=exc[j];
         for (j=0;j<nsf;j++)
            best_target[j]=new_target[j];
         best_err=err;
         best_pitch=pitch;
         best_gain_index=cdbk_index;
      }
   }
   
   /*printf ("pitch: %d %d\n", best_pitch, best_gain_index);*/
   speex_bits_pack(bits, best_pitch-start, params->pitch_bits);
   speex_bits_pack(bits, best_gain_index, params->gain_bits);
   /*printf ("encode pitch: %d %d\n", best_pitch, best_gain_index);*/
   for (i=0;i<nsf;i++)
      exc[i]=best_exc[i];
   for (i=0;i<nsf;i++)
      target[i]=best_target[i];

   return pitch;
}

void pitch_unquant_3tap(
spx_sig_t exc[],                    /* Excitation */
int   start,                    /* Smallest pitch value allowed */
int   end,                      /* Largest pitch value allowed */
spx_word16_t pitch_coef,               /* Voicing (pitch) coefficient */
const void *par,
int   nsf,                      /* Number of samples in subframe */
int *pitch_val,
spx_word16_t *gain_val,
SpeexBits *bits,
char *stack,
int count_lost,
int subframe_offset,
spx_word16_t last_pitch_gain,
int cdbk_offset
)
{
   int i;
   int pitch;
   int gain_index;
   spx_word16_t gain[3];
   const signed char *gain_cdbk;
   int gain_cdbk_size;
   const ltp_params *params;

   params = (const ltp_params*) par;
   gain_cdbk_size = 1<<params->gain_bits;
   gain_cdbk = params->gain_cdbk + 3*gain_cdbk_size*cdbk_offset;

   pitch = speex_bits_unpack_unsigned(bits, params->pitch_bits);
   pitch += start;
   gain_index = speex_bits_unpack_unsigned(bits, params->gain_bits);
   /*printf ("decode pitch: %d %d\n", pitch, gain_index);*/
#ifdef FIXED_POINT
   gain[0] = ADD16(32,(spx_word16_t)gain_cdbk[gain_index*3]);
   gain[1] = ADD16(32,(spx_word16_t)gain_cdbk[gain_index*3+1]);
   gain[2] = ADD16(32,(spx_word16_t)gain_cdbk[gain_index*3+2]);
#else
   gain[0] = 0.015625*gain_cdbk[gain_index*3]+.5;
   gain[1] = 0.015625*gain_cdbk[gain_index*3+1]+.5;
   gain[2] = 0.015625*gain_cdbk[gain_index*3+2]+.5;
#endif

   if (count_lost && pitch > subframe_offset)
   {
      spx_word16_t gain_sum;
      if (1) {
#ifdef FIXED_POINT
         spx_word16_t tmp = count_lost < 4 ? last_pitch_gain : SHR16(last_pitch_gain,1);
         if (tmp>62)
            tmp=62;
#else
         spx_word16_t tmp = count_lost < 4 ? last_pitch_gain : 0.5 * last_pitch_gain;
         if (tmp>.95)
            tmp=.95;
#endif
         gain_sum = gain_3tap_to_1tap(gain);

         if (gain_sum > tmp)
         {
            spx_word16_t fact = DIV32_16(SHL32(EXTEND32(tmp),14),gain_sum);
            for (i=0;i<3;i++)
               gain[i]=MULT16_16_Q14(fact,gain[i]);
         }

      }

   }

   *pitch_val = pitch;
   gain_val[0]=gain[0];
   gain_val[1]=gain[1];
   gain_val[2]=gain[2];

   {
      spx_sig_t *e[3];
      VARDECL(spx_sig_t *tmp2);
      ALLOC(tmp2, 3*nsf, spx_sig_t);
      e[0]=tmp2;
      e[1]=tmp2+nsf;
      e[2]=tmp2+2*nsf;
      
      for (i=0;i<3;i++)
      {
         int j;
         int pp=pitch+1-i;
#if 0
         for (j=0;j<nsf;j++)
         {
            if (j-pp<0)
               e[i][j]=exc[j-pp];
            else if (j-pp-pitch<0)
               e[i][j]=exc[j-pp-pitch];
            else
               e[i][j]=0;
         }
#else
         {
            int tmp1, tmp3;
            tmp1=nsf;
            if (tmp1>pp)
               tmp1=pp;
            for (j=0;j<tmp1;j++)
               e[i][j]=exc[j-pp];
            tmp3=nsf;
            if (tmp3>pp+pitch)
               tmp3=pp+pitch;
            for (j=tmp1;j<tmp3;j++)
               e[i][j]=exc[j-pp-pitch];
            for (j=tmp3;j<nsf;j++)
               e[i][j]=0;
         }
#endif
      }

#ifdef FIXED_POINT
      {
         for (i=0;i<nsf;i++)
            exc[i]=SHL32(ADD32(ADD32(MULT16_32_Q15(SHL16(gain[0],7),e[2][i]), MULT16_32_Q15(SHL16(gain[1],7),e[1][i])),
                               MULT16_32_Q15(SHL16(gain[2],7),e[0][i])), 2);
      }
#else
      for (i=0;i<nsf;i++)
         exc[i]=VERY_SMALL+gain[0]*e[2][i]+gain[1]*e[1][i]+gain[2]*e[0][i];
#endif
   }
}


/** Forced pitch delay and gain */
int forced_pitch_quant(
spx_sig_t target[],                 /* Target vector */
spx_sig_t *sw,
spx_coef_t ak[],                     /* LPCs for this subframe */
spx_coef_t awk1[],                   /* Weighted LPCs #1 for this subframe */
spx_coef_t awk2[],                   /* Weighted LPCs #2 for this subframe */
spx_sig_t exc[],                    /* Excitation */
const void *par,
int   start,                    /* Smallest pitch value allowed */
int   end,                      /* Largest pitch value allowed */
spx_word16_t pitch_coef,               /* Voicing (pitch) coefficient */
int   p,                        /* Number of LPC coeffs */
int   nsf,                      /* Number of samples in subframe */
SpeexBits *bits,
char *stack,
spx_sig_t *exc2,
spx_word16_t *r,
int complexity,
int cdbk_offset,
int plc_tuning
)
{
   int i;
   float coef = GAIN_SCALING_1*pitch_coef;
   if (coef>.99)
      coef=.99;
   for (i=0;i<nsf;i++)
   {
      exc[i]=exc[i-start]*coef;
   }
   return start;
}

/** Unquantize forced pitch delay and gain */
void forced_pitch_unquant(
spx_sig_t exc[],                    /* Excitation */
int   start,                    /* Smallest pitch value allowed */
int   end,                      /* Largest pitch value allowed */
spx_word16_t pitch_coef,               /* Voicing (pitch) coefficient */
const void *par,
int   nsf,                      /* Number of samples in subframe */
int *pitch_val,
spx_word16_t *gain_val,
SpeexBits *bits,
char *stack,
int count_lost,
int subframe_offset,
spx_word16_t last_pitch_gain,
int cdbk_offset
)
{
   int i;
   float coef = GAIN_SCALING_1*pitch_coef;
   if (coef>.99)
      coef=.99;
   for (i=0;i<nsf;i++)
   {
      exc[i]=exc[i-start]*coef;
   }
   *pitch_val = start;
   gain_val[0]=gain_val[2]=0;
   gain_val[1] = pitch_coef;
}