ref: 971b055090a8c553705acb6d853248545d56db4c
parent: 70d90d115d01cb47426295efca7a5a3ffcafd130
author: Jean-Marc Valin <[email protected]>
date: Fri Dec 2 11:08:02 EST 2011
New transient code, weighted tonality
--- a/celt/celt.c
+++ b/celt/celt.c
@@ -293,31 +293,32 @@
}
static int transient_analysis(const opus_val32 * restrict in, int len, int C,
- int overlap, opus_val16 *tf_estimate)
+ int overlap, opus_val16 *tf_estimate, int *tf_chan)
{int i;
VARDECL(opus_val16, tmp);
- opus_val32 mem0=0,mem1=0;
+ opus_val32 mem0,mem1;
int is_transient = 0;
int block;
- int N;
- opus_val16 maxbin;
- opus_val32 L1, L2, tf_tmp;
+ int c, N;
+ opus_val16 maxbin, minbin[3];
+ opus_val32 L1, L2, tf_tmp, tf_max;
VARDECL(opus_val16, bins);
SAVE_STACK;
ALLOC(tmp, len, opus_val16);
- block = overlap/2;
- N=len/block;
+ block = overlap/8;
+ N=len/block-1;
ALLOC(bins, N, opus_val16);
- if (C==1)
+
+ *tf_estimate = 0;
+ tf_max = 0;
+ for (c=0;c<C;c++)
{+ mem0=0;
+ mem1=0;
for (i=0;i<len;i++)
- tmp[i] = SHR32(in[i],SIG_SHIFT);
- } else {- for (i=0;i<len;i++)
- tmp[i] = SHR32(ADD32(in[i],in[i+len]), SIG_SHIFT+1);
- }
+ tmp[i] = SHR32(in[i*C+c],SIG_SHIFT);
/* High-pass filter: (1 - 2*z^-1 + z^-2) / (1 - z^-1 + .5*z^-2) */
for (i=0;i<len;i++)
@@ -339,15 +340,36 @@
tmp[i] = 0;
maxbin=0;
+ minbin[0] = minbin[1] = minbin[2] = 32768;
for (i=0;i<N;i++)
{int j;
opus_val16 max_abs=0;
- for (j=0;j<block;j++)
+ for (j=0;j<2*block;j++)
max_abs = MAX16(max_abs, ABS16(tmp[i*block+j]));
bins[i] = max_abs;
maxbin = MAX16(maxbin, bins[i]);
+ if (bins[i] < minbin[2])
+ {+ if (bins[i] < minbin[1])
+ {+ if (bins[i] < minbin[0])
+ {+ minbin[2] = minbin[1];
+ minbin[1] = minbin[0];
+ minbin[0] = bins[i];
+ } else {+ minbin[2] = minbin[1];
+ minbin[1] = bins[i];
+ }
+ } else {+ minbin[2] = bins[i];
+ }
+ }
}
+ //printf("%f ", maxbin/minbin[2]);+ if (maxbin > 15*minbin[2])
+ is_transient = 1;
L1=0;
L2=0;
for (i=0;i<N;i++)
@@ -361,7 +383,7 @@
L1 += EXTEND32(tmp_bin);
L2 += SHR32(MULT16_16(tmp_bin, tmp_bin), 4);
t1 = MULT16_16_Q15(QCONST16(.15f, 15), bins[i]);
- t2 = MULT16_16_Q15(QCONST16(.4f, 15), bins[i]);
+ t2 = MULT16_16_Q15(QCONST16(.3f, 15), bins[i]);
t3 = MULT16_16_Q15(QCONST16(.15f, 15), bins[i]);
for (j=0;j<i;j++)
{@@ -372,7 +394,7 @@
else
conseq = 0;
}
- if (conseq>=3)
+ if (conseq>=12)
is_transient=1;
conseq = 0;
for (j=i+1;j<N;j++)
@@ -382,16 +404,25 @@
else
conseq = 0;
}
- if (conseq>=7)
+ if (conseq>=28)
is_transient=1;
}
/* sqrt(L2*N)/L1 */
tf_tmp = SHL32(DIV32( SHL32(EXTEND32(celt_sqrt(SHR16(L2,4) * N)), 14), ADD32(EPSILON, L1)), 4);
- *tf_estimate = MAX16(QCONST16(1.f, 14), EXTRACT16(MIN16(QCONST32(1.99, 14), tf_tmp)));
+ tf_tmp = 1+MIN16(1,MAX16(0, 1-10*minbin[2]/(1+maxbin)));
+ if (tf_tmp>tf_max)
+ {+ *tf_chan = c;
+ tf_max = tf_tmp;
+ }
+ *tf_estimate = MAX16(*tf_estimate, EXTRACT16(MIN32(QCONST32(1.99, 14), tf_tmp)));
+ }
+ *tf_estimate = MAX16(QCONST16(1.f, 14), *tf_estimate);
RESTORE_STACK;
#ifdef FUZZING
is_transient = rand()&0x1;
#endif
+ //printf("%d %f\n", is_transient, *tf_estimate);return is_transient;
}
@@ -548,17 +579,15 @@
{0, -2, 0, -3, 3, 0, 1,-1},};
-static opus_val32 l1_metric(const celt_norm *tmp, int N, int LM)
+static opus_val32 l1_metric(const celt_norm *tmp, int N, int LM, opus_val16 bias)
{int i;
opus_val32 L1;
- opus_val16 bias;
L1 = 0;
for (i=0;i<N;i++)
L1 += EXTEND32(ABS16(tmp[i]));
/* When in doubt, prefer goo freq resolution */
- bias = QCONST16(.015f,15)*LM;
- L1 = MAC16_32_Q15(L1, bias, L1);
+ L1 = MAC16_32_Q15(L1, LM*bias, L1);
return L1;
}
@@ -565,7 +594,7 @@
static int tf_analysis(const CELTMode *m, int len, int C, int isTransient,
int *tf_res, int nbCompressedBytes, celt_norm *X, int N0, int LM,
- int *tf_sum)
+ int *tf_sum, opus_val16 tf_estimate, int tf_chan)
{int i;
VARDECL(int, metric);
@@ -576,7 +605,11 @@
VARDECL(celt_norm, tmp);
int lambda;
int tf_select=0;
+ opus_val16 bias;
+
SAVE_STACK;
+ bias = QCONST16(.04f,15)*MAX16(-.25, 1.5-tf_estimate);
+ /*printf("%f ", bias);*/if (nbCompressedBytes<15*C)
{@@ -607,12 +640,12 @@
int best_level=0;
N = (m->eBands[i+1]-m->eBands[i])<<LM;
for (j=0;j<N;j++)
- tmp[j] = X[j+(m->eBands[i]<<LM)];
+ tmp[j] = X[tf_chan*N0 + j+(m->eBands[i]<<LM)];
/* Just add the right channel if we're in stereo */
- if (C==2)
+ /*if (C==2)
for (j=0;j<N;j++)
- tmp[j] = ADD16(SHR16(tmp[j], 1),SHR16(X[N0+j+(m->eBands[i]<<LM)], 1));
- L1 = l1_metric(tmp, N, isTransient ? LM : 0);
+ tmp[j] = ADD16(SHR16(tmp[j], 1),SHR16(X[N0+j+(m->eBands[i]<<LM)], 1));*/
+ L1 = l1_metric(tmp, N, isTransient ? LM : 0, bias);
best_L1 = L1;
/*printf ("%f ", L1);*/for (k=0;k<LM;k++)
@@ -629,7 +662,7 @@
else
haar1(tmp, N>>k, 1<<k);
- L1 = l1_metric(tmp, N, B);
+ L1 = l1_metric(tmp, N, B, bias);
if (L1 < best_L1)
{@@ -642,7 +675,8 @@
metric[i] = best_level;
else
metric[i] = -best_level;
- *tf_sum += metric[i];
+ //printf("%d ", metric[i]);+ *tf_sum += (isTransient ? LM : 0) - metric[i];
}
/*printf("\n");*//* NOTE: Future optimized implementations could detect extreme transients and set
@@ -690,6 +724,7 @@
else
tf_res[i] = path0[i+1];
}
+ //printf("%d %f\n", *tf_sum, tf_estimate);RESTORE_STACK;
#ifdef FUZZING
tf_select = rand()&0x1;
@@ -737,7 +772,7 @@
tf_select = 0;
for (i=start;i<end;i++)
tf_res[i] = tf_select_table[LM][4*isTransient+2*tf_select+tf_res[i]];
- /*printf("%d %d ", isTransient, tf_select); for(i=0;i<end;i++)printf("%d ", tf_res[i]);printf("\n");*/+ //for(i=0;i<end;i++)printf("%d ", isTransient ? LM-tf_res[i] : -tf_res[i]);printf("\n");}
static void tf_decode(int start, int end, int isTransient, int *tf_res, int LM, ec_dec *dec)
@@ -957,6 +992,7 @@
int anti_collapse_rsv;
int anti_collapse_on=0;
int silence=0;
+ int tf_chan = 0;
opus_val16 tf_estimate=0;
opus_val16 stereo_saving = 0;
ALLOC_STACK;
@@ -1257,7 +1293,7 @@
if (st->complexity > 1)
{isTransient = transient_analysis(in, N+st->overlap, CC,
- st->overlap, &tf_estimate);
+ st->overlap, &tf_estimate, &tf_chan);
if (isTransient)
shortBlocks = M;
}
@@ -1291,12 +1327,15 @@
compute_band_energies(st->mode, freq, bandE, effEnd, C, M);
amp2Log2(st->mode, effEnd, st->end, bandE, bandLogE, C);
+ /*for (i=0;i<17;i++)
+ printf("%f ", bandLogE[i]);+ printf("\n");*//* Band normalisation */
normalise_bands(st->mode, freq, X, bandE, effEnd, C, M);
ALLOC(tf_res, st->mode->nbEBands, int);
- tf_select = tf_analysis(st->mode, effEnd, C, isTransient, tf_res, effectiveBytes, X, N, LM, &tf_sum);
+ tf_select = tf_analysis(st->mode, effEnd, C, isTransient, tf_res, effectiveBytes, X, N, LM, &tf_sum, tf_estimate, tf_chan);
for (i=effEnd;i<st->end;i++)
tf_res[i] = tf_res[effEnd-1];
@@ -1495,7 +1534,16 @@
#ifdef FIXED_POINT
new_target = SHL32(MULT16_32_Q15(target, SUB16(tf_estimate, QCONST16(0.05, 14))),1);
#else
- new_target = target*(tf_estimate-.05);
+ {+ //float tf_factor = 1+MIN16(1,2*MAX16(0,sqrt(tf_estimate-1)-.2));
+ float tf_factor = tf_estimate;
+ if (isTransient)
+ tf_factor = MAX16(1.2f, tf_factor);
+ //new_target = target*(tf_estimate-.05);
+ new_target = target*(tf_factor-.15);
+ //new_target = target*MIN32(2.f,MAX16(.85f,tf_sum/21.));
+ //printf("%f %f %f %f ", tf_factor, tf_sum/21., target*(tf_estimate-1.05), target*MIN32(2.f,MAX16(.85f,tf_sum/21.))-target);+ }
#endif
#ifndef FIXED_POINT
@@ -1502,11 +1550,12 @@
if (st->analysis.valid) {int tonal_target;
float tonal;
- tonal = st->analysis.tonality*st->analysis.tonality;
- tonal -= .08;
- tonal_target = target + (coded_bins<<BITRES)*1.5f*tonal;
+ tonal = st->analysis.tonality;
+ tonal -= .15;
+ tonal_target = target + (coded_bins<<BITRES)*1.6f*tonal;
/*printf("%f %d\n", tonal, tonal_target);*/new_target = IMAX(tonal_target,new_target);
+ //printf("%f %f ", tonal, (coded_bins<<BITRES)*1.6f*tonal);}
#endif
@@ -1569,7 +1618,7 @@
/*printf ("+%d\n", adjust);*/}
nbCompressedBytes = IMIN(nbCompressedBytes,nbAvailableBytes+nbFilledBytes);
- /*printf("%d\n", nbCompressedBytes*50*8);*/+ //printf("%d\n", nbCompressedBytes*50*8);/* This moves the raw bits to take into account the new compressed size */
ec_enc_shrink(enc, nbCompressedBytes);
}
--- a/src/analysis.c
+++ b/src/analysis.c
@@ -67,9 +67,13 @@
#define NB_TBANDS 18
static const int tbands[NB_TBANDS+1] = {- 2, 4, 6, 8, 10, 12, 14, 16, 20, 24, 28, 32, 40, 48, 56, 68, 80, 96, 120
+ 2, 4, 6, 8, 10, 12, 14, 16, 20, 24, 28, 32, 40, 48, 56, 68, 80, 96, 120
};
+static const float tweight[NB_TBANDS+1] = {+ .3, .4, .5, .6, .7, .8, .9, 1., 1., 1., 1., 1., 1., 1., .8, .7, .6, .5
+};
+
#define NB_TONAL_SKIP_BANDS 0
typedef struct {@@ -111,6 +115,8 @@
float BFCC[8];
float features[100];
float frame_tonality;
+ float max_frame_tonality;
+ float tw_sum=0;
float frame_noisiness;
const float pi4 = M_PI*M_PI*M_PI*M_PI;
float slope=0;
@@ -192,6 +198,8 @@
}
frame_tonality = 0;
+ max_frame_tonality = 0;
+ tw_sum = 0;
info->activity = 0;
frame_noisiness = 0;
frame_stationarity = 0;
@@ -257,8 +265,19 @@
frame_stationarity += stationarity;
/*band_tonality[b] = tE/(1e-15+E)*/;
band_tonality[b] = MAX16(tE/(EPSILON+E), stationarity*tonal->prev_band_tonality[b]);
+ //printf("%f ", band_tonality[b]);+#if 1
if (b>=NB_TONAL_SKIP_BANDS)
- frame_tonality += band_tonality[b];
+ {+ frame_tonality += tweight[b]*band_tonality[b];
+ tw_sum += tweight[b];
+ }
+#else
+ frame_tonality += band_tonality[b];
+ if (b>=NB_TBANDS-NB_TONAL_SKIP_BANDS)
+ frame_tonality -= band_tonality[b-NB_TBANDS+NB_TONAL_SKIP_BANDS];
+#endif
+ max_frame_tonality = MAX16(max_frame_tonality, frame_tonality);
slope += band_tonality[b]*(b-8);
/*printf("%f %f ", band_tonality[b], stationarity);*/if (band_tonality[b] > info->boost_amount[1] && b>=7 && b < NB_TBANDS-1)
@@ -276,7 +295,7 @@
}
tonal->prev_band_tonality[b] = band_tonality[b];
}
-
+ //printf("\n");frame_loudness = 20*log10(frame_loudness);
tonal->Etracker = MAX32(tonal->Etracker-.03, frame_loudness);
tonal->lowECount *= (1-alphaE);
@@ -301,7 +320,7 @@
#else
info->activity = .5*(1+frame_noisiness-frame_stationarity);
#endif
- frame_tonality /= NB_TBANDS-NB_TONAL_SKIP_BANDS;
+ frame_tonality = (max_frame_tonality/(tw_sum));
frame_tonality = MAX16(frame_tonality, tonal->prev_tonality*.8);
tonal->prev_tonality = frame_tonality;
info->boost_amount[0] -= frame_tonality+.2;