From 2824bd1f666102cff9e904ea8c6496b261960c03 Mon Sep 17 00:00:00 2001
From: Jean-Marc Valin <jmvalin@amazon.com>
Date: Wed, 28 Jun 2023 17:01:17 -0400
Subject: [PATCH] Adjusting offsets to synthesize 10 ms at a time
Should make synthesis easier in the future
---
dnn/lpcnet_plc.c | 10 +++++-----
dnn/lpcnet_private.h | 2 +-
silk/dred_config.h | 9 ++++++---
src/opus_decoder.c | 2 +-
src/opus_encoder.c | 2 +-
5 files changed, 14 insertions(+), 11 deletions(-)
diff --git a/dnn/lpcnet_plc.c b/dnn/lpcnet_plc.c
index d0067583f..879064aff 100644
--- a/dnn/lpcnet_plc.c
+++ b/dnn/lpcnet_plc.c
@@ -208,7 +208,7 @@ int lpcnet_plc_update(LPCNetPLCState *st, opus_int16 *pcm) {
pcm[i] = (int)floor(.5 + w*pcm[i] + (1-w)*(tmp[i]-delta));
}
st->lpcnet = copy;
- lpcnet_synthesize_impl(&st->lpcnet, &st->features[0], pcm, FRAME_SIZE-TRAINING_OFFSET, FRAME_SIZE-TRAINING_OFFSET);
+ /*lpcnet_synthesize_impl(&st->lpcnet, &st->features[0], pcm, FRAME_SIZE-TRAINING_OFFSET, FRAME_SIZE-TRAINING_OFFSET);*/
} else {
if (FEATURES_DELAY > 0) st->plc_net = st->plc_copy[FEATURES_DELAY-1];
fec_rewind(st, FEATURES_DELAY);
@@ -219,8 +219,8 @@ int lpcnet_plc_update(LPCNetPLCState *st, opus_int16 *pcm) {
lpcnet_synthesize_tail_impl(&st->lpcnet, tmp, FRAME_SIZE-TRAINING_OFFSET, FRAME_SIZE-TRAINING_OFFSET);
#endif
}
- OPUS_COPY(st->pcm, &pcm[FRAME_SIZE-TRAINING_OFFSET], TRAINING_OFFSET);
- st->pcm_fill = TRAINING_OFFSET;
+ OPUS_COPY(st->pcm, pcm, FRAME_SIZE);
+ st->pcm_fill = FRAME_SIZE;
} else {
OPUS_COPY(&st->pcm[st->pcm_fill], pcm, FRAME_SIZE);
st->pcm_fill += FRAME_SIZE;
@@ -286,12 +286,12 @@ int lpcnet_plc_conceal(LPCNetPLCState *st, opus_int16 *pcm) {
}
OPUS_MOVE(&st->plc_copy[1], &st->plc_copy[0], FEATURES_DELAY);
st->plc_copy[0] = st->plc_net;
- lpcnet_synthesize_tail_impl(&st->lpcnet, pcm, FRAME_SIZE-TRAINING_OFFSET, 0);
+ /*lpcnet_synthesize_tail_impl(&st->lpcnet, pcm, FRAME_SIZE-TRAINING_OFFSET, 0);*/
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));
else st->features[0] = MAX16(-10, st->features[0]+att_table[st->loss_count]);
- lpcnet_synthesize_impl(&st->lpcnet, &st->features[0], &pcm[FRAME_SIZE-TRAINING_OFFSET], TRAINING_OFFSET, 0);
+ lpcnet_synthesize_impl(&st->lpcnet, &st->features[0], pcm, FRAME_SIZE, 0);
{
float x[FRAME_SIZE];
/* FIXME: Can we do better? */
diff --git a/dnn/lpcnet_private.h b/dnn/lpcnet_private.h
index 87d142a14..da048e7e8 100644
--- a/dnn/lpcnet_private.h
+++ b/dnn/lpcnet_private.h
@@ -62,7 +62,7 @@ struct LPCNetEncState{
float burg_cepstrum[2*NB_BANDS];
};
-#define PLC_BUF_SIZE (FEATURES_DELAY*FRAME_SIZE + TRAINING_OFFSET)
+#define PLC_BUF_SIZE (FEATURES_DELAY*FRAME_SIZE + FRAME_SIZE)
struct LPCNetPLCState {
PLCModel model;
LPCNetState lpcnet;
diff --git a/silk/dred_config.h b/silk/dred_config.h
index ccaf76dbb..c190b5545 100644
--- a/silk/dred_config.h
+++ b/silk/dred_config.h
@@ -35,13 +35,16 @@
#define DRED_NUM_FEATURES 20
#define DRED_LATENT_DIM 80
#define DRED_STATE_DIM 24
-#define DRED_MAX_FRAMES 100
-#define DRED_SILK_ENCODER_DELAY (79+12)
+#define DRED_SILK_ENCODER_DELAY (79+12-80)
#define DRED_FRAME_SIZE 160
#define DRED_DFRAME_SIZE (2 * (DRED_FRAME_SIZE))
#define DRED_MAX_DATA_SIZE 1000
#define DRED_ENC_Q0 9
#define DRED_ENC_Q1 15
-#define DRED_NUM_REDUNDANCY_FRAMES 50
+
+/* Covers 1.04 second so we can cover one second, after the lookahead. */
+#define DRED_MAX_LATENTS 26
+#define DRED_NUM_REDUNDANCY_FRAMES (2*DRED_MAX_LATENTS)
+#define DRED_MAX_FRAMES (4*DRED_MAX_LATENTS)
#endif
diff --git a/src/opus_decoder.c b/src/opus_decoder.c
index 88571157d..c45facac4 100644
--- a/src/opus_decoder.c
+++ b/src/opus_decoder.c
@@ -673,7 +673,7 @@ int opus_decode_native(OpusDecoder *st, const unsigned char *data,
/* if blend==0, the last PLC call was "update" and we need to feed two extra 10-ms frames. */
if (st->lpcnet.blend == 0) needed_feature_frames+=2;
for (i=0;i<needed_feature_frames;i++) {
- int feature_offset = (needed_feature_frames-i-1 + (dred_offset/(st->Fs/100)-2));
+ int feature_offset = (needed_feature_frames-i-1 + (dred_offset/(st->Fs/100)-1));
if (feature_offset <= 4*dred->nb_latents-1 && feature_offset >= 0) {
lpcnet_plc_fec_add(&st->lpcnet, dred->fec_features+feature_offset*DRED_NUM_FEATURES);
} else {
diff --git a/src/opus_encoder.c b/src/opus_encoder.c
index 9f9f3fe33..b8e0f92aa 100644
--- a/src/opus_encoder.c
+++ b/src/opus_encoder.c
@@ -2205,7 +2205,7 @@ opus_int32 opus_encode_native(OpusEncoder *st, const opus_val16 *pcm, int frame_
unsigned char buf[DRED_MAX_DATA_SIZE];
int dred_chunks;
int dred_bytes_left;
- dred_chunks = IMIN(st->dred_duration/4, DRED_NUM_REDUNDANCY_FRAMES/2);
+ dred_chunks = IMIN((st->dred_duration+5)/4, DRED_NUM_REDUNDANCY_FRAMES/2);
dred_bytes_left = IMIN(DRED_MAX_DATA_SIZE, max_data_bytes-ret-2);
/* Check whether we actually have something to encode. */
if (dred_chunks >= 1 && dred_bytes_left >= DRED_MIN_BYTES+2) {
--
GitLab