shithub: opus

Info  •  Files  •  Log  •  Branches

reindent

--- a/dnn/dump_data.c

+++ b/dnn/dump_data.c

@@ -124,6 +124,10 @@

   float g;

   kiss_fft_cpx X[FREQ_SIZE];

   float Ex[NB_BANDS];

+  float xcorr[PITCH_MAX_PERIOD];

+  float ener0;

+  int sub;

+  float ener;

   RNN_COPY(aligned_in, &st->analysis_mem[OVERLAP_SIZE-TRAINING_OFFSET], TRAINING_OFFSET);

   frame_analysis(st, X, Ex, in);

   logMax = -2;

@@ -140,155 +144,148 @@

   g = lpc_from_cepstrum(st->lpc, st->features[st->pcount]);

   st->features[st->pcount][2*NB_BANDS+2] = log10(g);

   for (i=0;i<LPC_ORDER;i++) st->features[st->pcount][2*NB_BANDS+3+i] = st->lpc[i];

-  {

-    float xcorr[PITCH_MAX_PERIOD];

-    float ener0;

-    int sub;

-    float ener;

-    RNN_MOVE(st->exc_buf, &st->exc_buf[FRAME_SIZE], PITCH_MAX_PERIOD);

-    RNN_COPY(&aligned_in[TRAINING_OFFSET], in, FRAME_SIZE-TRAINING_OFFSET);

-    for (i=0;i<FRAME_SIZE;i++) {

-      int j;

-      float sum = aligned_in[i];

-      for (j=0;j<LPC_ORDER;j++)

-        sum += st->lpc[j]*st->pitch_mem[j];

-      RNN_MOVE(st->pitch_mem+1, st->pitch_mem, LPC_ORDER-1);

-      st->pitch_mem[0] = aligned_in[i];

-      st->exc_buf[PITCH_MAX_PERIOD+i] = sum + .7*st->pitch_filt;

-      st->pitch_filt = sum;

-      //printf("%f\n", st->exc_buf[PITCH_MAX_PERIOD+i]);

+  RNN_MOVE(st->exc_buf, &st->exc_buf[FRAME_SIZE], PITCH_MAX_PERIOD);

+  RNN_COPY(&aligned_in[TRAINING_OFFSET], in, FRAME_SIZE-TRAINING_OFFSET);

+  for (i=0;i<FRAME_SIZE;i++) {

+    int j;

+    float sum = aligned_in[i];

+    for (j=0;j<LPC_ORDER;j++)

+      sum += st->lpc[j]*st->pitch_mem[j];

+    RNN_MOVE(st->pitch_mem+1, st->pitch_mem, LPC_ORDER-1);

+    st->pitch_mem[0] = aligned_in[i];

+    st->exc_buf[PITCH_MAX_PERIOD+i] = sum + .7*st->pitch_filt;

+    st->pitch_filt = sum;

+    //printf("%f\n", st->exc_buf[PITCH_MAX_PERIOD+i]);

+  }

+  /* Cross-correlation on half-frames. */

+  for (sub=0;sub<2;sub++) {

+    int off = sub*FRAME_SIZE/2;

+    celt_pitch_xcorr(&st->exc_buf[PITCH_MAX_PERIOD+off], st->exc_buf+off, xcorr, FRAME_SIZE/2, PITCH_MAX_PERIOD);

+    ener0 = celt_inner_prod(&st->exc_buf[PITCH_MAX_PERIOD+off], &st->exc_buf[PITCH_MAX_PERIOD+off], FRAME_SIZE/2);

+    st->frame_weight[2+2*st->pcount+sub] = ener0;

+    //printf("%f\n", st->frame_weight[2+2*st->pcount+sub]);

+    for (i=0;i<PITCH_MAX_PERIOD;i++) {

+      ener = (1 + ener0 + celt_inner_prod(&st->exc_buf[i+off], &st->exc_buf[i+off], FRAME_SIZE/2));

+      st->xc[2+2*st->pcount+sub][i] = 2*xcorr[i] / ener;

     }

-    /* Cross-correlation on half-frames. */

-    for (sub=0;sub<2;sub++) {

-      int off = sub*FRAME_SIZE/2;

-      celt_pitch_xcorr(&st->exc_buf[PITCH_MAX_PERIOD+off], st->exc_buf+off, xcorr, FRAME_SIZE/2, PITCH_MAX_PERIOD);

-      ener0 = celt_inner_prod(&st->exc_buf[PITCH_MAX_PERIOD+off], &st->exc_buf[PITCH_MAX_PERIOD+off], FRAME_SIZE/2);

-      st->frame_weight[2+2*st->pcount+sub] = ener0;

-      //printf("%f\n", st->frame_weight[2+2*st->pcount+sub]);

-      for (i=0;i<PITCH_MAX_PERIOD;i++) {

-        ener = (1 + ener0 + celt_inner_prod(&st->exc_buf[i+off], &st->exc_buf[i+off], FRAME_SIZE/2));

-        st->xc[2+2*st->pcount+sub][i] = 2*xcorr[i] / ener;

-      }

 #if 0

-      for (i=0;i<PITCH_MAX_PERIOD;i++)

-        printf("%f ", st->xc[2*st->pcount+sub][i]);

-      printf("\n");

+    for (i=0;i<PITCH_MAX_PERIOD;i++)

+      printf("%f ", st->xc[2*st->pcount+sub][i]);

+    printf("\n");

 #endif

-    }

-    st->pcount++;

   }

+  st->pcount++;

 }

 static void process_superframe(DenoiseState *st, FILE *ffeat) {

   int i;

   int sub;

-      int best_i;

-      int best[10];

-      int pitch_prev[8][PITCH_MAX_PERIOD];

-      float best_a=0;

-      float best_b=0;

-      float w;

-      float sx=0, sxx=0, sxy=0, sy=0, sw=0;

-      float frame_corr;

-      int voiced;

-      float frame_weight_sum = 1e-15;

-      for(sub=0;sub<8;sub++) frame_weight_sum += st->frame_weight[2+sub];

-      for(sub=0;sub<8;sub++) st->frame_weight[2+sub] *= (8.f/frame_weight_sum);

-      for(sub=0;sub<8;sub++) {

-        float max_path_all = -1e15;

-        best_i = 0;

-        for (i=0;i<PITCH_MAX_PERIOD-2*PITCH_MIN_PERIOD;i++) {

-          float xc_half = MAX16(MAX16(st->xc[2+sub][(PITCH_MAX_PERIOD+i)/2], st->xc[2+sub][(PITCH_MAX_PERIOD+i+2)/2]), st->xc[2+sub][(PITCH_MAX_PERIOD+i-1)/2]);

-          if (st->xc[2+sub][i] < xc_half*1.1) st->xc[2+sub][i] *= .8;

+  int best_i;

+  int best[10];

+  int pitch_prev[8][PITCH_MAX_PERIOD];

+  float best_a=0;

+  float best_b=0;

+  float w;

+  float sx=0, sxx=0, sxy=0, sy=0, sw=0;

+  float frame_corr;

+  int voiced;

+  float frame_weight_sum = 1e-15;

+  for(sub=0;sub<8;sub++) frame_weight_sum += st->frame_weight[2+sub];

+  for(sub=0;sub<8;sub++) st->frame_weight[2+sub] *= (8.f/frame_weight_sum);

+  for(sub=0;sub<8;sub++) {

+    float max_path_all = -1e15;

+    best_i = 0;

+    for (i=0;i<PITCH_MAX_PERIOD-2*PITCH_MIN_PERIOD;i++) {

+      float xc_half = MAX16(MAX16(st->xc[2+sub][(PITCH_MAX_PERIOD+i)/2], st->xc[2+sub][(PITCH_MAX_PERIOD+i+2)/2]), st->xc[2+sub][(PITCH_MAX_PERIOD+i-1)/2]);

+      if (st->xc[2+sub][i] < xc_half*1.1) st->xc[2+sub][i] *= .8;

+    }

+    for (i=0;i<PITCH_MAX_PERIOD-PITCH_MIN_PERIOD;i++) {

+      int j;

+      float max_prev;

+      max_prev = st->pitch_max_path_all - 6.f;

+      pitch_prev[sub][i] = st->best_i;

+      for (j=IMIN(0, 4-i);j<=4 && i+j<PITCH_MAX_PERIOD-PITCH_MIN_PERIOD;j++) {

+        if (st->pitch_max_path[0][i+j] > max_prev) {

+          max_prev = st->pitch_max_path[0][i+j] - .02f*abs(j)*abs(j);

+          pitch_prev[sub][i] = i+j;

         }

-        for (i=0;i<PITCH_MAX_PERIOD-PITCH_MIN_PERIOD;i++) {

-          int j;

-          float max_prev;

-          max_prev = st->pitch_max_path_all - 6.f;

-          pitch_prev[sub][i] = st->best_i;

-          for (j=IMIN(0, 4-i);j<=4 && i+j<PITCH_MAX_PERIOD-PITCH_MIN_PERIOD;j++) {

-            if (st->pitch_max_path[0][i+j] > max_prev) {

-              max_prev = st->pitch_max_path[0][i+j] - .02f*abs(j)*abs(j);

-              pitch_prev[sub][i] = i+j;

-            }

-          }

-          st->pitch_max_path[1][i] = max_prev + st->frame_weight[2+sub]*st->xc[2+sub][i];

-          if (st->pitch_max_path[1][i] > max_path_all) {

-            max_path_all = st->pitch_max_path[1][i];

-            best_i = i;

-          }

-        }

-        /* Renormalize. */

-        for (i=0;i<PITCH_MAX_PERIOD-PITCH_MIN_PERIOD;i++) st->pitch_max_path[1][i] -= max_path_all;

-        //for (i=0;i<PITCH_MAX_PERIOD-PITCH_MIN_PERIOD;i++) printf("%f ", st->pitch_max_path[1][i]);

-        //printf("\n");

-        RNN_COPY(&st->pitch_max_path[0][0], &st->pitch_max_path[1][0], PITCH_MAX_PERIOD);

-        st->pitch_max_path_all = max_path_all;

-        st->best_i = best_i;

       }

-      best_i = st->best_i;

-      frame_corr = 0;

-      /* Backward pass. */

-      for (sub=7;sub>=0;sub--) {

-        best[2+sub] = PITCH_MAX_PERIOD-best_i;

-        frame_corr += st->frame_weight[2+sub]*st->xc[2+sub][best_i];

-        best_i = pitch_prev[sub][best_i];

+      st->pitch_max_path[1][i] = max_prev + st->frame_weight[2+sub]*st->xc[2+sub][i];

+      if (st->pitch_max_path[1][i] > max_path_all) {

+        max_path_all = st->pitch_max_path[1][i];

+        best_i = i;

       }

-      frame_corr /= 8;

-      for (sub=0;sub<8;sub++) {

-        //printf("%d %f\n", best[2+sub], frame_corr);

-      }

-      //printf("\n");

-      for (sub=2;sub<10;sub++) {

-        w = st->frame_weight[sub];

-        sw += w;

-        sx += w*sub;

-        sxx += w*sub*sub;

-        sxy += w*sub*best[sub];

-        sy += w*best[sub];

-      }

-      voiced = frame_corr > .3;

-      /* Linear regression to figure out the pitch contour. */

-      best_a = (sw*sxy - sx*sy)/(sw*sxx - sx*sx);

-      if (voiced) {

-        float mean_pitch = sy/sw;

-        /* Allow a relative variation of up to 1/4 over 8 sub-frames. */

-        float max_a = mean_pitch/32;

-        best_a = MIN16(max_a, MAX16(-max_a, best_a));

-      } else {

-        best_a = 0;

-      }

-      //best_b = (sxx*sy - sx*sxy)/(sw*sxx - sx*sx);

-      best_b = (sy - best_a*sx)/sw;

-      /* Quantizing the pitch as "main" pitch + slope. */

-      float center_pitch = best_b+5.5*best_a;

-      int main_pitch = (int)floor(.5 + 21.*log2(center_pitch/PITCH_MIN_PERIOD));

-      main_pitch = IMAX(0, IMIN(63, main_pitch));

-      int modulation = (int)floor(.5 + 16*7*best_a/center_pitch);

-      modulation = IMAX(-3, IMIN(3, modulation));

-      //printf("%d %d\n", main_pitch, modulation);

-      //printf("%f %f\n", best_a/center_pitch, best_corr);

-      //for (sub=2;sub<10;sub++) printf("%f %d %f\n", best_b + sub*best_a, best[sub], best_corr);

-      for (sub=0;sub<4;sub++) {

+    }

+    /* Renormalize. */

+    for (i=0;i<PITCH_MAX_PERIOD-PITCH_MIN_PERIOD;i++) st->pitch_max_path[1][i] -= max_path_all;

+    //for (i=0;i<PITCH_MAX_PERIOD-PITCH_MIN_PERIOD;i++) printf("%f ", st->pitch_max_path[1][i]);

+    //printf("\n");

+    RNN_COPY(&st->pitch_max_path[0][0], &st->pitch_max_path[1][0], PITCH_MAX_PERIOD);

+    st->pitch_max_path_all = max_path_all;

+    st->best_i = best_i;

+  }

+  best_i = st->best_i;

+  frame_corr = 0;

+  /* Backward pass. */

+  for (sub=7;sub>=0;sub--) {

+    best[2+sub] = PITCH_MAX_PERIOD-best_i;

+    frame_corr += st->frame_weight[2+sub]*st->xc[2+sub][best_i];

+    best_i = pitch_prev[sub][best_i];

+  }

+  frame_corr /= 8;

+  for (sub=0;sub<8;sub++) {

+    //printf("%d %f\n", best[2+sub], frame_corr);

+  }

+  //printf("\n");

+  for (sub=2;sub<10;sub++) {

+    w = st->frame_weight[sub];

+    sw += w;

+    sx += w*sub;

+    sxx += w*sub*sub;

+    sxy += w*sub*best[sub];

+    sy += w*best[sub];

+  }

+  voiced = frame_corr > .3;

+  /* Linear regression to figure out the pitch contour. */

+  best_a = (sw*sxy - sx*sy)/(sw*sxx - sx*sx);

+  if (voiced) {

+    float mean_pitch = sy/sw;

+    /* Allow a relative variation of up to 1/4 over 8 sub-frames. */

+    float max_a = mean_pitch/32;

+    best_a = MIN16(max_a, MAX16(-max_a, best_a));

+  } else {

+    best_a = 0;

+  }

+  //best_b = (sxx*sy - sx*sxy)/(sw*sxx - sx*sx);

+  best_b = (sy - best_a*sx)/sw;

+  /* Quantizing the pitch as "main" pitch + slope. */

+  float center_pitch = best_b+5.5*best_a;

+  int main_pitch = (int)floor(.5 + 21.*log2(center_pitch/PITCH_MIN_PERIOD));

+  main_pitch = IMAX(0, IMIN(63, main_pitch));

+  int modulation = (int)floor(.5 + 16*7*best_a/center_pitch);

+  modulation = IMAX(-3, IMIN(3, modulation));

+  //printf("%d %d\n", main_pitch, modulation);

+  //printf("%f %f\n", best_a/center_pitch, best_corr);

+  //for (sub=2;sub<10;sub++) printf("%f %d %f\n", best_b + sub*best_a, best[sub], best_corr);

+  for (sub=0;sub<4;sub++) {

 #if 0

-          float p = pow(2.f, main_pitch/21.)*PITCH_MIN_PERIOD;

-          p *= 1 + modulation/16./7.*(2*sub-3);

-          st->features[sub][2*NB_BANDS] = .02*(p-100);

-          st->features[sub][2*NB_BANDS + 1] = voiced ? 1 : -1;

+    float p = pow(2.f, main_pitch/21.)*PITCH_MIN_PERIOD;

+    p *= 1 + modulation/16./7.*(2*sub-3);

+    st->features[sub][2*NB_BANDS] = .02*(p-100);

+    st->features[sub][2*NB_BANDS + 1] = voiced ? 1 : -1;

 #else

-          st->features[sub][2*NB_BANDS] = .01*(best[2+2*sub]+best[2+2*sub+1]-200);

-          st->features[sub][2*NB_BANDS + 1] = frame_corr-.5;

+    st->features[sub][2*NB_BANDS] = .01*(best[2+2*sub]+best[2+2*sub+1]-200);

+    st->features[sub][2*NB_BANDS + 1] = frame_corr-.5;

 #endif

-          //printf("%f %d %f %f\n", st->features[sub][2*NB_BANDS], best[2+2*sub], best_corr, frame_corr);

-      }

-      //printf("%d %f %f %f\n", best_period, best_a, best_b, best_corr);

-      RNN_COPY(&st->xc[0][0], &st->xc[8][0], PITCH_MAX_PERIOD);

-      RNN_COPY(&st->xc[1][0], &st->xc[9][0], PITCH_MAX_PERIOD);

-

-      for (i=0;i<4;i++) {

-        fwrite(st->features[i], sizeof(float), NB_FEATURES, ffeat);

-      }

-      st->pcount=0;

+    //printf("%f %d %f %f\n", st->features[sub][2*NB_BANDS], best[2+2*sub], best_corr, frame_corr);

+  }

+  //printf("%d %f %f %f\n", best_period, best_a, best_b, best_corr);

+  RNN_COPY(&st->xc[0][0], &st->xc[8][0], PITCH_MAX_PERIOD);

+  RNN_COPY(&st->xc[1][0], &st->xc[9][0], PITCH_MAX_PERIOD);

+  for (i=0;i<4;i++) {

+    fwrite(st->features[i], sizeof(float), NB_FEATURES, ffeat);

+  }

+  st->pcount=0;

 }

 static void biquad(float *y, float mem[2], const float *x, const float *b, const float *a, int N) {

-- 

⑨