nb_celp.c

speex 1.1.12 编码 稳定版本

#ifdef EPIC_48K
      if (st->lbr_48k)
      {
         float ol1=0,ol2=0;
         float ol_gain2;
         ol1 = compute_rms(st->exc, st->frameSize>>1);
         ol2 = compute_rms(st->exc+(st->frameSize>>1), st->frameSize>>1);
         ol1 *= ol1*(st->frameSize>>1);
         ol2 *= ol2*(st->frameSize>>1);

         ol_gain2=ol1;
         if (ol2>ol1)
            ol_gain2=ol2;
         ol_gain2 = sqrt(2*ol_gain2*(ol1+ol2))*1.3*(1-.5*GAIN_SCALING_1*GAIN_SCALING_1*ol_pitch_coef*ol_pitch_coef);
      
         ol_gain=SHR(sqrt(1+ol_gain2/st->frameSize),SIG_SHIFT);

      } else {
#endif
         ol_gain = SHL32(EXTEND32(compute_rms(st->exc, st->frameSize)),SIG_SHIFT);
#ifdef EPIC_48K
      }
#endif
   }

#ifdef VORBIS_PSYCHO
   compute_curve(st->psy, st->frame-16, st->curve);
   /*print_vec(st->curve, 128, "curve");*/
   if (st->first)
      for (i=0;i<128;i++)
         st->old_curve[i] = st->curve[i];
#endif

   /*VBR stuff*/
   if (st->vbr && (st->vbr_enabled||st->vad_enabled))
   {
      float lsp_dist=0;
      for (i=0;i<st->lpcSize;i++)
         lsp_dist += (st->old_lsp[i] - st->lsp[i])*(st->old_lsp[i] - st->lsp[i]);
      lsp_dist /= LSP_SCALING*LSP_SCALING;
      
      if (st->abr_enabled)
      {
         float qual_change=0;
         if (st->abr_drift2 * st->abr_drift > 0)
         {
            /* Only adapt if long-term and short-term drift are the same sign */
            qual_change = -.00001*st->abr_drift/(1+st->abr_count);
            if (qual_change>.05)
               qual_change=.05;
            if (qual_change<-.05)
               qual_change=-.05;
         }
         st->vbr_quality += qual_change;
         if (st->vbr_quality>10)
            st->vbr_quality=10;
         if (st->vbr_quality<0)
            st->vbr_quality=0;
      }

      st->relative_quality = vbr_analysis(st->vbr, in, st->frameSize, ol_pitch, GAIN_SCALING_1*ol_pitch_coef);
      /*if (delta_qual<0)*/
      /*  delta_qual*=.1*(3+st->vbr_quality);*/
      if (st->vbr_enabled) 
      {
         int mode;
         int choice=0;
         float min_diff=100;
         mode = 8;
         while (mode)
         {
            int v1;
            float thresh;
            v1=(int)floor(st->vbr_quality);
            if (v1==10)
               thresh = vbr_nb_thresh[mode][v1];
            else
               thresh = (st->vbr_quality-v1)*vbr_nb_thresh[mode][v1+1] + (1+v1-st->vbr_quality)*vbr_nb_thresh[mode][v1];
            if (st->relative_quality > thresh && 
                st->relative_quality-thresh<min_diff)
            {
               choice = mode;
               min_diff = st->relative_quality-thresh;
            }
            mode--;
         }
         mode=choice;
         if (mode==0)
         {
            if (st->dtx_count==0 || lsp_dist>.05 || !st->dtx_enabled || st->dtx_count>20)
            {
               mode=1;
               st->dtx_count=1;
            } else {
               mode=0;
               st->dtx_count++;
            }
         } else {
            st->dtx_count=0;
         }

         speex_encoder_ctl(state, SPEEX_SET_MODE, &mode);

         if (st->abr_enabled)
         {
            int bitrate;
            speex_encoder_ctl(state, SPEEX_GET_BITRATE, &bitrate);
            st->abr_drift+=(bitrate-st->abr_enabled);
            st->abr_drift2 = .95*st->abr_drift2 + .05*(bitrate-st->abr_enabled);
            st->abr_count += 1.0;
         }

      } else {
         /*VAD only case*/
         int mode;
         if (st->relative_quality<2)
         {
            if (st->dtx_count==0 || lsp_dist>.05 || !st->dtx_enabled || st->dtx_count>20)
            {
               st->dtx_count=1;
               mode=1;
            } else {
               mode=0;
               st->dtx_count++;
            }
         } else {
            st->dtx_count = 0;
            mode=st->submodeSelect;
         }
         /*speex_encoder_ctl(state, SPEEX_SET_MODE, &mode);*/
         st->submodeID=mode;
      } 
   } else {
      st->relative_quality = -1;
   }

   if (st->encode_submode)
   {
#ifdef EPIC_48K
   if (!st->lbr_48k) {
#endif

   /* First, transmit a zero for narrowband */
   speex_bits_pack(bits, 0, 1);

   /* Transmit the sub-mode we use for this frame */
   speex_bits_pack(bits, st->submodeID, NB_SUBMODE_BITS);

#ifdef EPIC_48K
   }
#endif
   }

   /* If null mode (no transmission), just set a couple things to zero*/
   if (st->submodes[st->submodeID] == NULL)
   {
      for (i=0;i<st->frameSize;i++)
         st->exc[i]=st->sw[i]=VERY_SMALL;

      for (i=0;i<st->lpcSize;i++)
         st->mem_sw[i]=0;
      st->first=1;
      st->bounded_pitch = 1;

      /* Final signal synthesis from excitation */
      iir_mem2(st->exc, st->interp_qlpc, st->frame, st->frameSize, st->lpcSize, st->mem_sp);

#ifdef RESYNTH
      for (i=0;i<st->frameSize;i++)
         in[i]=st->frame[i];
#endif
      return 0;

   }

   /* LSP Quantization */
   if (st->first)
   {
      for (i=0;i<st->lpcSize;i++)
         st->old_lsp[i] = st->lsp[i];
   }


   /*Quantize LSPs*/
#if 1 /*0 for unquantized*/
   SUBMODE(lsp_quant)(st->lsp, st->qlsp, st->lpcSize, bits);
#else
   for (i=0;i<st->lpcSize;i++)
     st->qlsp[i]=st->lsp[i];
#endif

#ifdef EPIC_48K
   if (st->lbr_48k) {
      speex_bits_pack(bits, pitch_half[0]-st->min_pitch, 7);
      speex_bits_pack(bits, pitch_half[1]-pitch_half[0]+1, 2);
      
      {
         int quant = (int)floor(.5+7.4*GAIN_SCALING_1*ol_pitch_coef);
         if (quant>7)
            quant=7;
         if (quant<0)
            quant=0;
         ol_pitch_id=quant;
         speex_bits_pack(bits, quant, 3);
         ol_pitch_coef=GAIN_SCALING*0.13514*quant;
         
      }
      {
         int qe = (int)(floor(.5+2.1*log(ol_gain*1.0/SIG_SCALING)))-2;
         if (qe<0)
            qe=0;
         if (qe>15)
            qe=15;
         ol_gain = exp((qe+2)/2.1)*SIG_SCALING;
         speex_bits_pack(bits, qe, 4);
      }

   } else {
#endif

   /*If we use low bit-rate pitch mode, transmit open-loop pitch*/
   if (SUBMODE(lbr_pitch)!=-1)
   {
      speex_bits_pack(bits, ol_pitch-st->min_pitch, 7);
   } 

   if (SUBMODE(forced_pitch_gain))
   {
      int quant;
      quant = (int)floor(.5+15*ol_pitch_coef*GAIN_SCALING_1);
      if (quant>15)
         quant=15;
      if (quant<0)
         quant=0;
      speex_bits_pack(bits, quant, 4);
      ol_pitch_coef=GAIN_SCALING*0.066667*quant;
   }
   
   
   /*Quantize and transmit open-loop excitation gain*/
#ifdef FIXED_POINT
   {
      int qe = scal_quant32(ol_gain, ol_gain_table, 32);
      /*ol_gain = exp(qe/3.5)*SIG_SCALING;*/
      ol_gain = MULT16_32_Q15(28406,ol_gain_table[qe]);
      speex_bits_pack(bits, qe, 5);
   }
#else
   {
      int qe = (int)(floor(.5+3.5*log(ol_gain*1.0/SIG_SCALING)));
      if (qe<0)
         qe=0;
      if (qe>31)
         qe=31;
      ol_gain = exp(qe/3.5)*SIG_SCALING;
      speex_bits_pack(bits, qe, 5);
   }
#endif


#ifdef EPIC_48K
   }
#endif


   /* Special case for first frame */
   if (st->first)
   {
      for (i=0;i<st->lpcSize;i++)
         st->old_qlsp[i] = st->qlsp[i];
   }

   /* Filter response */
   ALLOC(res, st->subframeSize, spx_sig_t);
   /* Target signal */
   ALLOC(target, st->subframeSize, spx_sig_t);
   ALLOC(syn_resp, st->subframeSize, spx_word16_t);
   ALLOC(real_exc, st->subframeSize, spx_sig_t);
   ALLOC(mem, st->lpcSize, spx_mem_t);

   /* Loop on sub-frames */
   for (sub=0;sub<st->nbSubframes;sub++)
   {
      int   offset;
      spx_sig_t *sp, *sw, *exc;
      int pitch;
      int response_bound = st->subframeSize;
#ifdef EPIC_48K
      if (st->lbr_48k)
      {
         if (sub*2 < st->nbSubframes)
            ol_pitch = pitch_half[0];
         else
            ol_pitch = pitch_half[1];
      }
#endif

      /* Offset relative to start of frame */
      offset = st->subframeSize*sub;
      /* Original signal */
      sp=st->frame+offset;
      /* Excitation */
      exc=st->exc+offset;
      /* Weighted signal */
      sw=st->sw+offset;

      /* LSP interpolation (quantized and unquantized) */
      lsp_interpolate(st->old_lsp, st->lsp, st->interp_lsp, st->lpcSize, sub, st->nbSubframes);
      lsp_interpolate(st->old_qlsp, st->qlsp, st->interp_qlsp, st->lpcSize, sub, st->nbSubframes);

      /* Make sure the filters are stable */
      lsp_enforce_margin(st->interp_lsp, st->lpcSize, LSP_MARGIN);
      lsp_enforce_margin(st->interp_qlsp, st->lpcSize, LSP_MARGIN);

      /* Compute interpolated LPCs (quantized and unquantized) */
      lsp_to_lpc(st->interp_lsp, st->interp_lpc, st->lpcSize,stack);

      lsp_to_lpc(st->interp_qlsp, st->interp_qlpc, st->lpcSize, stack);

      /* Compute analysis filter gain at w=pi (for use in SB-CELP) */
      {
         spx_word32_t pi_g=LPC_SCALING;
         for (i=0;i<st->lpcSize;i+=2)
         {
            /*pi_g += -st->interp_qlpc[i] +  st->interp_qlpc[i+1];*/
            pi_g = ADD32(pi_g, SUB32(st->interp_qlpc[i+1],st->interp_qlpc[i]));
         }
         st->pi_gain[sub] = pi_g;
      }

#ifdef VORBIS_PSYCHO
      {
         float curr_curve[128];
         float fact = ((float)sub+1.0f)/st->nbSubframes;
         for (i=0;i<128;i++)
            curr_curve[i] = (1.0f-fact)*st->old_curve[i] + fact*st->curve[i];
         curve_to_lpc(st->psy, curr_curve, st->bw_lpc1, st->bw_lpc2, 10);
      }
#else
      /* Compute bandwidth-expanded (unquantized) LPCs for perceptual weighting */
      bw_lpc(st->gamma1, st->interp_lpc, st->bw_lpc1, st->lpcSize);
      if (st->gamma2>=0)
         bw_lpc(st->gamma2, st->interp_lpc, st->bw_lpc2, st->lpcSize);
      else
      {
         st->bw_lpc2[0]=1;
         for (i=1;i<=st->lpcSize;i++)
            st->bw_lpc2[i]=0;
      }
      /*print_vec(st->bw_lpc1, 10, "bw_lpc");*/
#endif

      for (i=0;i<st->subframeSize;i++)
         real_exc[i] = exc[i];
      
      if (st->complexity==0)
         response_bound >>= 1;
      compute_impulse_response(st->interp_qlpc, st->bw_lpc1, st->bw_lpc2, syn_resp, response_bound, st->lpcSize, stack);
      for (i=response_bound;i<st->subframeSize;i++)
         syn_resp[i]=VERY_SMALL;
      
      /* Reset excitation */
      for (i=0;i<st->subframeSize;i++)
         exc[i]=VERY_SMALL;

      /* Compute zero response of A(z/g1) / ( A(z/g2) * A(z) ) */
      for (i=0;i<st->lpcSize;i++)
         mem[i]=st->mem_sp[i];
#ifdef SHORTCUTS2
      iir_mem2(exc, st->interp_qlpc, exc, response_bound, st->lpcSize, mem);
      for (i=0;i<st->lpcSize;i++)
         mem[i]=st->mem_sw[i];
      filter_mem2(exc, st->bw_lpc1, st->bw_lpc2, res, response_bound, st->lpcSize, mem);
      for (i=response_bound;i<st->subframeSize;i++)
         res[i]=0;
#else
      iir_mem2(exc, st->interp_qlpc, exc, st->subframeSize, st->lpcSize, mem);
      for (i=0;i<st->lpcSize;i++)
         mem[i]=st->mem_sw[i];
      filter_mem2(exc, st->bw_lpc1, st->bw_lpc2, res, st->subframeSize, st->lpcSize, mem);
#endif
      
      /* Compute weighted signal */
      for (i=0;i<st->lpcSize;i++)
         mem[i]=st->mem_sw[i];
      filter_mem2(sp, st->bw_lpc1, st->bw_lpc2, sw, st->subframeSize, st->lpcSize, mem);
      
      if (st->complexity==0)
         for (i=0;i<st->lpcSize;i++)
            st->mem_sw[i]=mem[i];
      
      /* Compute target signal */
      for (i=0;i<st->subframeSize;i++)
         target[i]=sw[i]-res[i];

      for (i=0;i<st->subframeSize;i++)
         exc[i]=0;

      /* If we have a long-term predictor (otherwise, something's wrong) */
      if (SUBMODE(ltp_quant))
      {
         int pit_min, pit_max;
         /* Long-term prediction */
         if (SUBMODE(lbr_pitch) != -1)
         {
            /* Low bit-rate pitch handling */
            int margin;
            margin = SUBMODE(lbr_pitch);
            if (margin)
            {
               if (ol_pitch < st->min_pitch+margin-1)
                  ol_pitch=st->min_pitch+margin-1;
               if (ol_pitch > st->max_pitch-margin)
                  ol_pitch=st->max_pitch-margin;
               pit_min = ol_pitch-margin+1;
               pit_max = ol_pitch+margin;
            } else {
               pit_min=pit_max=ol_pitch;