From c1b80a7ccffefde806a0674a19c756e2099db6b0 Mon Sep 17 00:00:00 2001
From: Jean-Marc Valin <jmvalin@jmvalin.ca>
Date: Mon, 15 Jan 2024 18:24:15 -0500
Subject: Improving PLC

Should handle the history in a more consistent way.
Slightly increase the model size and re-enable biased band loss in
training.
---
 autogen.sh           |  2 +-
 dnn/lpcnet_plc.c     | 17 +++++++++++++++--
 dnn/lpcnet_private.h |  4 +++-
 3 files changed, 19 insertions(+), 4 deletions(-)

diff --git a/autogen.sh b/autogen.sh
index 2cac2083..a961a9f8 100755
--- a/autogen.sh
+++ b/autogen.sh
@@ -9,7 +9,7 @@ set -e
 srcdir=`dirname $0`
 test -n "$srcdir" && cd "$srcdir"
 
-dnn/download_model.sh caca188
+dnn/download_model.sh 26ddfd7
 
 echo "Updating build configuration files, please wait...."
 
diff --git a/dnn/lpcnet_plc.c b/dnn/lpcnet_plc.c
index ec8ad7e9..5d397dd8 100644
--- a/dnn/lpcnet_plc.c
+++ b/dnn/lpcnet_plc.c
@@ -51,6 +51,8 @@ void lpcnet_plc_reset(LPCNetPLCState *st) {
   st->blend = 0;
   st->loss_count = 0;
   st->analysis_gap = 1;
+  st->analysis_pos = PLC_BUF_SIZE;
+  st->predict_pos = PLC_BUF_SIZE;
 }
 
 int lpcnet_plc_init(LPCNetPLCState *st) {
@@ -58,7 +60,6 @@ int lpcnet_plc_init(LPCNetPLCState *st) {
   st->arch = opus_select_arch();
   fargan_init(&st->fargan);
   lpcnet_encoder_init(&st->enc);
-  st->analysis_pos = PLC_BUF_SIZE;
   st->loaded = 0;
 #ifndef USE_WEIGHTS_FILE
   ret = init_plc_model(&st->model, lpcnet_plc_arrays);
@@ -148,6 +149,7 @@ int lpcnet_plc_update(LPCNetPLCState *st, opus_int16 *pcm) {
   int i;
   if (st->analysis_pos - FRAME_SIZE >= 0) st->analysis_pos -= FRAME_SIZE;
   else st->analysis_gap = 1;
+  if (st->predict_pos - FRAME_SIZE >= 0) st->predict_pos -= FRAME_SIZE;
   OPUS_MOVE(st->pcm, &st->pcm[FRAME_SIZE], PLC_BUF_SIZE-FRAME_SIZE);
   for (i=0;i<FRAME_SIZE;i++) st->pcm[PLC_BUF_SIZE-FRAME_SIZE+i] = (1.f/32768.f)*pcm[i];
   st->loss_count = 0;
@@ -161,28 +163,38 @@ int lpcnet_plc_conceal(LPCNetPLCState *st, opus_int16 *pcm) {
   celt_assert(st->loaded);
   if (st->blend == 0) {
     int count = 0;
+    st->plc_net = st->plc_bak[0];
     while (st->analysis_pos + FRAME_SIZE <= PLC_BUF_SIZE) {
       float x[FRAME_SIZE];
       float plc_features[2*NB_BANDS+NB_FEATURES+1];
+      celt_assert(st->analysis_pos >= 0);
       for (i=0;i<FRAME_SIZE;i++) x[i] = 32768.f*st->pcm[st->analysis_pos+i];
       burg_cepstral_analysis(plc_features, x);
       lpcnet_compute_single_frame_features_float(&st->enc, x, st->features, st->arch);
-      if ((st->analysis_gap && count > 0) || count > 1) {
+      if ((!st->analysis_gap || count>0) && st->analysis_pos + 0*FRAME_SIZE >= st->predict_pos) {
         queue_features(st, st->features);
         OPUS_COPY(&plc_features[2*NB_BANDS], st->features, NB_FEATURES);
         plc_features[2*NB_BANDS+NB_FEATURES] = 1;
+        st->plc_bak[0] = st->plc_bak[1];
+        st->plc_bak[1] = st->plc_net;
         compute_plc_pred(st, st->features, plc_features);
       }
       st->analysis_pos += FRAME_SIZE;
       count++;
     }
+    st->plc_bak[0] = st->plc_bak[1];
+    st->plc_bak[1] = st->plc_net;
     get_fec_or_pred(st, st->features);
     queue_features(st, st->features);
+    st->plc_bak[0] = st->plc_bak[1];
+    st->plc_bak[1] = st->plc_net;
     get_fec_or_pred(st, st->features);
     queue_features(st, st->features);
     fargan_cont(&st->fargan, &st->pcm[PLC_BUF_SIZE-FARGAN_CONT_SAMPLES], st->cont_features);
     st->analysis_gap = 0;
   }
+  st->plc_bak[0] = st->plc_bak[1];
+  st->plc_bak[1] = st->plc_net;
   if (get_fec_or_pred(st, st->features)) st->loss_count = 0;
   else st->loss_count++;
   if (st->loss_count >= 10) st->features[0] = MAX16(-10, st->features[0]+att_table[9] - 2*(st->loss_count-9));
@@ -191,6 +203,7 @@ int lpcnet_plc_conceal(LPCNetPLCState *st, opus_int16 *pcm) {
   queue_features(st, st->features);
   if (st->analysis_pos - FRAME_SIZE >= 0) st->analysis_pos -= FRAME_SIZE;
   else st->analysis_gap = 1;
+  st->predict_pos = PLC_BUF_SIZE;
   OPUS_MOVE(st->pcm, &st->pcm[FRAME_SIZE], PLC_BUF_SIZE-FRAME_SIZE);
   for (i=0;i<FRAME_SIZE;i++) st->pcm[PLC_BUF_SIZE-FRAME_SIZE+i] = (1.f/32768.f)*pcm[i];
   st->blend = 1;
diff --git a/dnn/lpcnet_private.h b/dnn/lpcnet_private.h
index 7fb8123a..6ff3ddb2 100644
--- a/dnn/lpcnet_private.h
+++ b/dnn/lpcnet_private.h
@@ -41,7 +41,7 @@ struct LPCNetEncState{
   float burg_cepstrum[2*NB_BANDS];
 };
 
-#define PLC_BUF_SIZE (CONT_VECTORS*FRAME_SIZE)
+#define PLC_BUF_SIZE ((CONT_VECTORS+5)*FRAME_SIZE)
 struct LPCNetPLCState {
   PLCModel model;
   FARGANState fargan;
@@ -56,12 +56,14 @@ struct LPCNetPLCState {
   int fec_fill_pos;
   int fec_skip;
   int analysis_pos;
+  int predict_pos;
   float pcm[PLC_BUF_SIZE];
   int blend;
   float features[NB_TOTAL_FEATURES];
   float cont_features[CONT_VECTORS*NB_FEATURES];
   int loss_count;
   PLCNetState plc_net;
+  PLCNetState plc_bak[2];
 };
 
 void preemphasis(float *y, float *mem, const float *x, float coef, int N);
-- 
cgit v1.2.3