shithub: opus

Info  •  Files  •  Log  •  Branches

Fix quantization bug where pitch can get too low

Would cause unused pitch embedding vectors to be used

--- a/dnn/lpcnet.c

+++ b/dnn/lpcnet.c

@@ -130,6 +130,7 @@

     float pitch_gain;

     /* Matches the Python code -- the 0.1 avoids rounding issues. */

     pitch = (int)floor(.1 + 50*features[36]+100);

+    pitch = IMIN(255, IMAX(33, pitch));

     pitch_gain = lpcnet->old_gain[FEATURES_DELAY-1];

     memmove(&lpcnet->old_gain[1], &lpcnet->old_gain[0], (FEATURES_DELAY-1)*sizeof(lpcnet->old_gain[0]));

     lpcnet->old_gain[0] = features[PITCH_GAIN_FEATURE];

--- a/dnn/lpcnet_dec.c

+++ b/dnn/lpcnet_dec.c

@@ -123,6 +123,7 @@

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

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

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

+    p = MIN16(255, MAX16(33, p));

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

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

   }

--- a/dnn/lpcnet_enc.c

+++ b/dnn/lpcnet_enc.c

@@ -662,11 +662,11 @@

     if (quantize) {

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

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

-      p = MIN16(255, MAX16(32, p));

+      p = MIN16(255, MAX16(33, p));

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

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

     } else {

-      st->features[sub][2*NB_BANDS] = .01*(IMAX(64, IMIN(510, best[2+2*sub]+best[2+2*sub+1]))-200);

+      st->features[sub][2*NB_BANDS] = .01*(IMAX(66, IMIN(510, best[2+2*sub]+best[2+2*sub+1]))-200);

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

     }

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

-- 

⑨