sb_celp.c

mediastreamer2是开源的网络传输媒体流的库

   ALLOC(interp_lpc, st->lpcSize, spx_coef_t);
   ALLOC(bw_lpc1, st->lpcSize, spx_coef_t);
   ALLOC(bw_lpc2, st->lpcSize, spx_coef_t);
   
   ALLOC(lsp, st->lpcSize, spx_lsp_t);
   ALLOC(qlsp, st->lpcSize, spx_lsp_t);
   ALLOC(interp_lsp, st->lpcSize, spx_lsp_t);
   ALLOC(interp_qlsp, st->lpcSize, spx_lsp_t);
   
   {
      VARDECL(spx_word16_t *autocorr);
      VARDECL(spx_word16_t *w_sig);
      ALLOC(autocorr, st->lpcSize+1, spx_word16_t);
      ALLOC(w_sig, st->windowSize, spx_word16_t);
      /* Window for analysis */
      /* FIXME: This is a kludge */
      if (st->subframeSize==80)
      {
         for (i=0;i<st->windowSize;i++)
            w_sig[i] = EXTRACT16(SHR32(MULT16_16(high[i],st->window[i>>1]),SIG_SHIFT));
      } else {
         for (i=0;i<st->windowSize;i++)
            w_sig[i] = EXTRACT16(SHR32(MULT16_16(high[i],st->window[i]),SIG_SHIFT));
      }
      /* Compute auto-correlation */
      _spx_autocorr(w_sig, autocorr, st->lpcSize+1, st->windowSize);
      autocorr[0] = ADD16(autocorr[0],MULT16_16_Q15(autocorr[0],st->lpc_floor)); /* Noise floor in auto-correlation domain */

      /* Lag windowing: equivalent to filtering in the power-spectrum domain */
      for (i=0;i<st->lpcSize+1;i++)
         autocorr[i] = MULT16_16_Q14(autocorr[i],st->lagWindow[i]);

      /* Levinson-Durbin */
      _spx_lpc(lpc, autocorr, st->lpcSize);
   }

   /* LPC to LSPs (x-domain) transform */
   roots=lpc_to_lsp (lpc, st->lpcSize, lsp, 10, LSP_DELTA1, stack);
   if (roots!=st->lpcSize)
   {
      roots = lpc_to_lsp (lpc, st->lpcSize, lsp, 10, LSP_DELTA2, stack);
      if (roots!=st->lpcSize) {
         /*If we can't find all LSP's, do some damage control and use a flat filter*/
         for (i=0;i<st->lpcSize;i++)
         {
            lsp[i]=st->old_lsp[i];
         }
      }
   }

#ifndef DISABLE_VBR
   /* VBR code */
   if ((st->vbr_enabled || st->vad_enabled) && !dtx)
   {
      float ratio;
      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>.1)
               qual_change=.1;
            if (qual_change<-.1)
               qual_change=-.1;
         }
         st->vbr_quality += qual_change;
         if (st->vbr_quality>10)
            st->vbr_quality=10;
         if (st->vbr_quality<0)
            st->vbr_quality=0;
      }


      ratio = 2*log((1.f+e_high)/(1.f+e_low));
      
      speex_encoder_ctl(st->st_low, SPEEX_GET_RELATIVE_QUALITY, &st->relative_quality);
      if (ratio<-4)
         ratio=-4;
      if (ratio>2)
         ratio=2;
      /*if (ratio>-2)*/
      if (st->vbr_enabled) 
      {
         spx_int32_t modeid;
         modeid = mode->nb_modes-1;
         st->relative_quality+=1.0*(ratio+2);
	 if (st->relative_quality<-1)
            st->relative_quality=-1;
         while (modeid)
         {
            int v1;
            float thresh;
            v1=(int)floor(st->vbr_quality);
            if (v1==10)
               thresh = mode->vbr_thresh[modeid][v1];
            else
               thresh = (st->vbr_quality-v1)   * mode->vbr_thresh[modeid][v1+1] + 
                        (1+v1-st->vbr_quality) * mode->vbr_thresh[modeid][v1];
            if (st->relative_quality >= thresh && st->sampling_rate*st->submodes[modeid]->bits_per_frame/st->full_frame_size <= st->vbr_max_high)
               break;
            modeid--;
         }
         speex_encoder_ctl(state, SPEEX_SET_HIGH_MODE, &modeid);
         if (st->abr_enabled)
         {
            spx_int32_t 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 */
         int modeid;
         if (st->relative_quality<2.0)
            modeid=1;
         else
            modeid=st->submodeSelect;
         /*speex_encoder_ctl(state, SPEEX_SET_MODE, &mode);*/
         st->submodeID=modeid;

      }
      /*fprintf (stderr, "%f %f\n", ratio, low_qual);*/
   }
#endif /* #ifndef DISABLE_VBR */

   if (st->encode_submode)
   {
      speex_bits_pack(bits, 1, 1);
      if (dtx)
         speex_bits_pack(bits, 0, SB_SUBMODE_BITS);
      else
         speex_bits_pack(bits, st->submodeID, SB_SUBMODE_BITS);
   }

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

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

      /* Final signal synthesis from excitation */
      iir_mem16(high, st->interp_qlpc, high, st->frame_size, st->lpcSize, st->mem_sp, stack);

      if (dtx)
         return 0;
      else
         return 1;
   }


   /* LSP quantization */
   SUBMODE(lsp_quant)(lsp, qlsp, st->lpcSize, bits);   

   if (st->first)
   {
      for (i=0;i<st->lpcSize;i++)
         st->old_lsp[i] = lsp[i];
      for (i=0;i<st->lpcSize;i++)
         st->old_qlsp[i] = qlsp[i];
   }
   
   ALLOC(mem, st->lpcSize, spx_mem_t);
   ALLOC(syn_resp, st->subframeSize, spx_word16_t);
   ALLOC(innov, st->subframeSize, spx_sig_t);
   ALLOC(target, st->subframeSize, spx_word16_t);

   for (sub=0;sub<st->nbSubframes;sub++)
   {
      VARDECL(spx_word16_t *exc);
      VARDECL(spx_word16_t *res);
      VARDECL(spx_word16_t *sw);
      spx_word16_t *sp;
      spx_word16_t filter_ratio;     /*Q7*/
      int offset;
      spx_word32_t rl, rh;           /*Q13*/
      spx_word16_t eh=0;

      offset = st->subframeSize*sub;
      sp=high+offset;
      ALLOC(exc, st->subframeSize, spx_word16_t);
      ALLOC(res, st->subframeSize, spx_word16_t);
      ALLOC(sw, st->subframeSize, spx_word16_t);
      
      /* LSP interpolation (quantized and unquantized) */
      lsp_interpolate(st->old_lsp, lsp, interp_lsp, st->lpcSize, sub, st->nbSubframes);
      lsp_interpolate(st->old_qlsp, qlsp, interp_qlsp, st->lpcSize, sub, st->nbSubframes);

      lsp_enforce_margin(interp_lsp, st->lpcSize, LSP_MARGIN);
      lsp_enforce_margin(interp_qlsp, st->lpcSize, LSP_MARGIN);

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

      bw_lpc(st->gamma1, interp_lpc, bw_lpc1, st->lpcSize);
      bw_lpc(st->gamma2, interp_lpc, bw_lpc2, st->lpcSize);

      /* Compute mid-band (4000 Hz for wideband) response of low-band and high-band
         filters */
      st->pi_gain[sub]=LPC_SCALING;
      rh = LPC_SCALING;
      for (i=0;i<st->lpcSize;i+=2)
      {
         rh += st->interp_qlpc[i+1] - st->interp_qlpc[i];
         st->pi_gain[sub] += st->interp_qlpc[i] + st->interp_qlpc[i+1];
      }
      
      rl = low_pi_gain[sub];
#ifdef FIXED_POINT
      filter_ratio=EXTRACT16(SATURATE(PDIV32(SHL32(ADD32(rl,82),7),ADD32(82,rh)),32767));
#else
      filter_ratio=(rl+.01)/(rh+.01);
#endif
      
      /* Compute "real excitation" */
      fir_mem16(sp, st->interp_qlpc, exc, st->subframeSize, st->lpcSize, st->mem_sp2, stack);
      /* Compute energy of low-band and high-band excitation */

      eh = compute_rms16(exc, st->subframeSize);

      if (!SUBMODE(innovation_quant)) {/* 1 for spectral folding excitation, 0 for stochastic */
         spx_word32_t g;   /*Q7*/
         spx_word16_t el;  /*Q0*/
         el = low_innov_rms[sub];

         /* Gain to use if we want to use the low-band excitation for high-band */
         g=PDIV32(MULT16_16(filter_ratio,eh),EXTEND32(ADD16(1,el)));
         
#if 0
         {
            char *tmp_stack=stack;
            float *tmp_sig;
            float g2;
            ALLOC(tmp_sig, st->subframeSize, spx_sig_t);
            for (i=0;i<st->lpcSize;i++)
               mem[i]=st->mem_sp[i];
            iir_mem2(st->low_innov+offset, st->interp_qlpc, tmp_sig, st->subframeSize, st->lpcSize, mem);
            g2 = compute_rms(sp, st->subframeSize)/(.01+compute_rms(tmp_sig, st->subframeSize));
            /*fprintf (stderr, "gains: %f %f\n", g, g2);*/
            g = g2;
            stack = tmp_stack;
         }
#endif

         /*print_vec(&g, 1, "gain factor");*/
         /* Gain quantization */
         {
            int quant = scal_quant(g, fold_quant_bound, 32);
            /*speex_warning_int("tata", quant);*/
            if (quant<0)
               quant=0;
            if (quant>31)
               quant=31;
            speex_bits_pack(bits, quant, 5);
         }
         if (st->innov_rms_save)
         {
            st->innov_rms_save[sub] = eh;
         }
         st->exc_rms[sub] = eh;
      } else {
         spx_word16_t gc;       /*Q7*/
         spx_word32_t scale;    /*Q14*/
         spx_word16_t el;       /*Q0*/
         el = low_exc_rms[sub]; /*Q0*/

         gc = PDIV32_16(MULT16_16(filter_ratio,1+eh),1+el);

         /* This is a kludge that cleans up a historical bug */
         if (st->subframeSize==80)
            gc = MULT16_16_P15(QCONST16(0.70711f,15),gc);
         /*printf ("%f %f %f %f\n", el, eh, filter_ratio, gc);*/
         {
            int qgc = scal_quant(gc, gc_quant_bound, 16);
            speex_bits_pack(bits, qgc, 4);
            gc = MULT16_16_Q15(QCONST16(0.87360,15),gc_quant_bound[qgc]);
         }
         if (st->subframeSize==80)
            gc = MULT16_16_P14(QCONST16(1.4142f,14), gc);

         scale = SHL32(MULT16_16(PDIV32_16(SHL32(EXTEND32(gc),SIG_SHIFT-6),filter_ratio),(1+el)),6);

         compute_impulse_response(st->interp_qlpc, bw_lpc1, bw_lpc2, syn_resp, st->subframeSize, st->lpcSize, stack);

         
         /* Reset excitation */
         for (i=0;i<st->subframeSize;i++)
            res[i]=VERY_SMALL;
         
         /* Compute zero response (ringing) of A(z/g1) / ( A(z/g2) * Aq(z) ) */
         for (i=0;i<st->lpcSize;i++)
            mem[i]=st->mem_sp[i];
         iir_mem16(res, st->interp_qlpc, res, st->subframeSize, st->lpcSize, mem, stack);

         for (i=0;i<st->lpcSize;i++)
            mem[i]=st->mem_sw[i];
         filter_mem16(res, bw_lpc1, bw_lpc2, res, st->subframeSize, st->lpcSize, mem, stack);

         /* Compute weighted signal */
         for (i=0;i<st->lpcSize;i++)
            mem[i]=st->mem_sw[i];
         filter_mem16(sp, bw_lpc1, bw_lpc2, sw, st->subframeSize, st->lpcSize, mem, stack);

         /* Compute target signal */
         for (i=0;i<st->subframeSize;i++)
            target[i]=SUB16(sw[i],res[i]);

         signal_div(target, target, scale, st->subframeSize);

         /* Reset excitation */
         SPEEX_MEMSET(innov, 0, st->subframeSize);

         /*print_vec(target, st->subframeSize, "\ntarget");*/
         SUBMODE(innovation_quant)(target, st->interp_qlpc, bw_lpc1, bw_lpc2, 
                                   SUBMODE(innovation_params), st->lpcSize, st->subframeSize, 
                                   innov, syn_resp, bits, stack, st->complexity, SUBMODE(double_codebook));
         /*print_vec(target, st->subframeSize, "after");*/

         signal_mul(innov, innov, scale, st->subframeSize);

         if (SUBMODE(double_codebook)) {
            char *tmp_stack=stack;
            VARDECL(spx_sig_t *innov2);
            ALLOC(innov2, st->subframeSize, spx_sig_t);
            SPEEX_MEMSET(innov2, 0, st->subframeSize);
            for (i=0;i<st->subframeSize;i++)
               target[i]=MULT16_16_P13(QCONST16(2.5f,13), target[i]);

            SUBMODE(innovation_quant)(target, st->interp_qlpc, bw_lpc1, bw_lpc2, 
                                      SUBMODE(innovation_params), st->lpcSize, st->subframeSize, 
                                      innov2, syn_resp, bits, stack, st->complexity, 0);
            signal_mul(innov2, innov2, MULT16_32_P15(QCONST16(0.4f,15),scale), st->subframeSize);

            for (i=0;i<st->subframeSize;i++)
               innov[i] = ADD32(innov[i],innov2[i]);
            stack = tmp_stack;
         }
         for (i=0;i<st->subframeSize;i++)
            exc[i] = PSHR32(innov[i],SIG_SHIFT);

         if (st->innov_rms_save)
         {
            st->innov_rms_save[sub] = MULT16_16_Q15(QCONST16(.70711f, 15), compute_rms(innov, st->subframeSize));
         }
         st->exc_rms[sub] = compute_rms16(exc, st->subframeSize);
         

      }

      
      /*Keep the previous memory*/
      for (i=0;i<st->lpcSize;i++)
         mem[i]=st->mem_sp[i];
      /* Final signal synthesis from excitation */
      iir_mem16(exc, st->interp_qlpc, sp, st->subframeSize, st->lpcSize, st->mem_sp, stack);
      
      /* Compute weighted signal again, from synthesized speech (not sure it's the right thing) */
      filter_mem16(sp, bw_lpc1, bw_lpc2, sw, st->subframeSize, st->lpcSize, st->mem_sw, stack);
   }

   for (i=0;i<st->lpcSize;i++)
      st->old_lsp[i] = lsp[i];
   for (i=0;i<st->lpcSize;i++)
      st->old_qlsp[i] = qlsp[i];

   st->first=0;