********************************************************************
function: PCM data envelope analysis
- last mod: $Id: envelope.c,v 1.44 2002/03/24 21:04:00 xiphmont Exp $
+ last mod: $Id: envelope.c,v 1.45 2002/03/29 07:10:39 xiphmont Exp $
********************************************************************/
codec_setup_info *ci=vi->codec_setup;
vorbis_info_psy_global *gi=&ci->psy_g_param;
int ch=vi->channels;
- int i;
- int n=e->winlength=ci->blocksizes[0];
- e->searchstep=ci->blocksizes[0]/VE_DIV; /* not random */
+ int i,j;
+ int n=e->winlength=128;
+ e->searchstep=64; /* not random */
e->minenergy=gi->preecho_minenergy;
e->ch=ch;
mdct_init(&e->mdct,n);
for(i=0;i<n;i++){
- e->mdct_win[i]=sin((i+.5)/n*M_PI);
+ e->mdct_win[i]=sin(i/(n-1.)*M_PI);
e->mdct_win[i]*=e->mdct_win[i];
}
- /* overlapping bands, assuming 22050 (which is not always true, but
- to Hell with that) */
- /* 2(1.3-3) 4(2.6-6) 8(5.3-12) 16(10.6-18) */
-
- e->band[0].begin=rint(3000.f/22050.f*n/4.f)*2.f;
- e->band[0].end=rint(6000.f/22050.f*n/4.f)*2.f-e->band[0].begin;
- e->band[1].begin=rint(5000.f/22050.f*n/4.f)*2.f;
- e->band[1].end=rint(10000.f/22050.f*n/4.f)*2.f-e->band[1].begin;
- e->band[2].begin=rint(8000.f/22050.f*n/4.f)*2.f;
- e->band[2].end=rint(16000.f/22050.f*n/4.f)*2.f-e->band[2].begin;
- e->band[3].begin=rint(12000.f/22050.f*n/4.f)*2.f;
- e->band[3].end=rint(20000.f/22050.f*n/4.f)*2.f-e->band[3].begin;
-
- e->band[0].window=_ogg_malloc((e->band[0].end)/2*sizeof(*e->band[0].window));
- e->band[1].window=_ogg_malloc((e->band[1].end)/2*sizeof(*e->band[1].window));
- e->band[2].window=_ogg_malloc((e->band[2].end)/2*sizeof(*e->band[2].window));
- e->band[3].window=_ogg_malloc((e->band[3].end)/2*sizeof(*e->band[3].window));
-
- n=e->band[0].end/2;
- for(i=0;i<n;i++){
- e->band[0].window[i]=sin((i+.5)/n*M_PI);
- e->band[0].total+=e->band[0].window[i];
- }
- n=e->band[1].end/2;
- for(i=0;i<n;i++){
- e->band[1].window[i]=sin((i+.5)/n*M_PI);
- e->band[1].total+=e->band[1].window[i];
- }
- n=e->band[2].end/2;
- for(i=0;i<n;i++){
- e->band[2].window[i]=sin((i+.5)/n*M_PI);
- e->band[2].total+=e->band[2].window[i];
- }
- n=e->band[3].end/2;
- for(i=0;i<n;i++){
- e->band[3].window[i]=sin((i+.5)/n*M_PI);
- e->band[3].total+=e->band[3].window[i];
- }
+ /* magic follows */
+ e->band[0].begin=4; e->band[0].end=8;
+ e->band[1].begin=8; e->band[1].end=10;
+ e->band[2].begin=14; e->band[2].end=12;
+ e->band[3].begin=20; e->band[3].end=16;
+ e->band[4].begin=28; e->band[4].end=20;
+ e->band[5].begin=40; e->band[5].end=20;
+ for(j=0;j<VE_BANDS;j++){
+ n=e->band[j].end;
+ e->band[j].window=_ogg_malloc(n*sizeof(*e->band[0].window));
+ for(i=0;i<n;i++){
+ e->band[j].window[i]=sin((i+.5)/n*M_PI);
+ e->band[j].total+=e->band[j].window[i];
+ }
+ e->band[j].total=1./e->band[j].total;
+ }
e->filter=_ogg_calloc(VE_BANDS*ch,sizeof(*e->filter));
e->mark=_ogg_calloc(e->storage,sizeof(*e->mark));
-
}
void _ve_envelope_clear(envelope_lookup *e){
memset(e,0,sizeof(*e));
}
+extern void _analysis_output_always(char *base,int i,float *v,int n,int bark,int dB,ogg_int64_t off);
+
/* fairly straight threshhold-by-band based until we find something
that works better and isn't patented. */
+
static int _ve_amp(envelope_lookup *ve,
vorbis_info_psy_global *gi,
float *data,
long n=ve->winlength;
int ret=0;
long i,j;
+ float decay;
/* we want to have a 'minimum bar' for energy, else we're just
basing blocks on quantization noise that outweighs the signal
itself (for low power signals) */
- float minV=ve->minenergy,acc[VE_BANDS];
+ float minV=ve->minenergy;
float *vec=alloca(n*sizeof(*vec));
- memset(acc,0,sizeof(acc));
-
+
+ /* stretch is used to gradually lengthen the number of windows
+ considered prevoius-to-potential-trigger */
+ int stretch=max(VE_MINSTRETCH,ve->stretch/2);
+ float penalty=gi->stretch_penalty-(ve->stretch/2-VE_MINSTRETCH);
+ if(penalty<0.f)penalty=0.f;
+ if(penalty>gi->stretch_penalty)penalty=gi->stretch_penalty;
+
+ /*_analysis_output_always("lpcm",seq2,data,n,0,0,
+ totalshift+pos*ve->searchstep);*/
+
/* window and transform */
for(i=0;i<n;i++)
vec[i]=data[i]*ve->mdct_win[i];
mdct_forward(&ve->mdct,vec,vec);
+
+ /* _analysis_output_always("mdct",seq2,vec,n/2,0,1,0); */
- /* accumulate amplitude by band */
- for(j=0;j<VE_BANDS;j++){
- for(i=0;i<bands[j].end;i+=2){
- float val=FABS(vec+i+bands[j].begin)+FABS(vec+i+bands[j].begin+1);
- acc[j]+=todB(&val)*bands[j].window[i>>1];
+ /* near-DC spreading function; this has nothing to do with
+ psychoacoustics, just sidelobe leakage and window size */
+ {
+ float temp=vec[0]*vec[0]+.7*vec[1]*vec[1]+.2*vec[2]*vec[2];
+ int ptr=filters->nearptr;
+
+ /* the accumulation is regularly refreshed from scratch to avoid
+ floating point creep */
+ if(ptr==0){
+ decay=filters->nearDC_acc=filters->nearDC_partialacc+temp;
+ filters->nearDC_partialacc=temp;
+ }else{
+ decay=filters->nearDC_acc+=temp;
+ filters->nearDC_partialacc+=temp;
}
- acc[j]/=bands[j].total;
- if(acc[j]<minV)acc[j]=minV;
- //fprintf(stderr,"%d %gdB :: ",j,acc[j]);
- }
- //fprintf(stderr,"\n");
+ filters->nearDC_acc-=filters->nearDC[ptr];
+ filters->nearDC[ptr]=temp;
- /* convert amplitude to delta */
- for(j=0;j<VE_BANDS;j++){
- float val=acc[j]-filters[j].ampbuf[filters[j].ampptr];
- filters[j].ampbuf[filters[j].ampptr]=acc[j];
- acc[j]=val;
- filters[j].ampptr++;
- if(filters[j].ampptr>=VE_DIV)filters[j].ampptr=0;
+ decay*=(1./(VE_NEARDC+1));
+ filters->nearptr++;
+ if(filters->nearptr>=VE_NEARDC)filters->nearptr=0;
+ decay=todB(&decay)*.5-15.f;
}
-
- /* convolve deltas to threshhold values */
- for(j=0;j<VE_BANDS;j++){
- float *buf=filters[j].delbuf;
- float val=.14*buf[0]+.14*buf[1]+.72*acc[j];
- buf[0]=buf[1];buf[1]=acc[j];
- acc[j]=val;
- /*filters[j].markers[pos+1]=val;*/
+
+ /* perform spreading and limiting, also smooth the spectrum. yes,
+ the MDCT results in all real coefficients, but it still *behaves*
+ like real/imaginary pairs */
+ for(i=0;i<n/2;i+=2){
+ float val=vec[i]*vec[i]+vec[i+1]*vec[i+1];
+ val=todB(&val)*.5f;
+ if(val<decay)val=decay;
+ if(val<minV)val=minV;
+ vec[i>>1]=val;
+ decay-=8.;
}
- /* look at local min/max */
+ /* _analysis_output_always("spread",seq2++,vec,n/4,0,0,0);*/
+
+ /* perform preecho/postecho triggering by band */
for(j=0;j<VE_BANDS;j++){
- float *buf=filters[j].convbuf;
- if(buf[1]>gi->preecho_thresh[j] && buf[0]<buf[1] && acc[j]<buf[1])ret|=1;
- if(buf[1]<gi->postecho_thresh[j] && buf[0]>buf[1] && acc[j]>buf[1])ret|=2;
- buf[0]=buf[1];buf[1]=acc[j];
+ float acc=0.;
+ float valmax,valmin;
+
+ /* accumulate amplitude */
+ for(i=0;i<bands[j].end;i++)
+ acc+=vec[i+bands[j].begin]*bands[j].window[i];
+
+ acc*=bands[j].total;
+
+ /* convert amplitude to delta */
+ {
+ int p,this=filters[j].ampptr;
+ float postmax,postmin,premax=-99999.f,premin=99999.f;
+
+ p=this;
+ p--;
+ if(p<0)p+=VE_AMP;
+ postmax=max(acc,filters[j].ampbuf[p]);
+ postmin=min(acc,filters[j].ampbuf[p]);
+
+ for(i=0;i<stretch;i++){
+ p--;
+ if(p<0)p+=VE_AMP;
+ premax=max(premax,filters[j].ampbuf[p]);
+ premin=min(premin,filters[j].ampbuf[p]);
+ }
+
+ valmin=postmin-premin;
+ valmax=postmax-premax;
+
+ filters[j].ampbuf[this]=acc;
+ filters[j].ampptr++;
+ if(filters[j].ampptr>=VE_AMP)filters[j].ampptr=0;
+ }
+
+ /* look at min/max, decide trigger */
+ if(valmax>gi->preecho_thresh[j]+penalty)ret|=1;
+ if(valmin<gi->postecho_thresh[j]-penalty)ret|=2;
}
+
+ if(ret&1)ve->stretch=-1;
+
return(ret);
}
long i,j;
int first=ve->current/ve->searchstep;
- int last=v->pcm_current/ve->searchstep-VE_DIV;
+ int last=v->pcm_current/ve->searchstep-VE_WIN;
if(first<0)first=0;
/* make sure we have enough storage to match the PCM */
if(last>ve->storage){
- ve->storage=last+VE_DIV;
+ ve->storage=last+VE_WIN;
ve->mark=_ogg_realloc(ve->mark,ve->storage*sizeof(*ve->mark));
}
for(j=first;j<last;j++){
int ret=0;
+
+ ve->stretch++;
+ if(ve->stretch>VE_MAXSTRETCH*2)
+ ve->stretch=VE_MAXSTRETCH*2;
+
for(i=0;i<ve->ch;i++){
- /* the mark delay is one searchstep because of min/max finder */
- float *pcm=v->pcm[i]+ve->searchstep*(j+1);
+ float *pcm=v->pcm[i]+ve->searchstep*(j);
ret|=_ve_amp(ve,gi,pcm,ve->band,ve->filter+i*VE_BANDS,j);
}
- /* we assume a 'transient' occupies half a short block; this way,
- it's contained in two short blocks, else the first block is
- short and the second long, causing smearing.
-
- preecho triggers follow the impulse marker; postecho triger preceed it */
+ ve->mark[j+VE_POST]=0;
+ if(ret&1){
+ ve->mark[j]=1;
+ ve->mark[j+1]=1;
+ }
- ve->mark[j+VE_DIV/2]=0;
- if(ret&1)
- for(i=0;i<=VE_DIV/2;i++)
- ve->mark[j+i]=1;
- if(ret&2)
- for(i=-1;i>=-VE_DIV/2 && j+i>=0;i--)
- ve->mark[j+i]=1;
+ if(ret&2){
+ ve->mark[j]=1;
+ ve->mark[j-1]=1;
+ }
}
ve->current=last*ve->searchstep;
j=ve->cursor;
- while(j<ve->current-(VE_DIV/2*ve->searchstep)){ /* modified to
- stay clear of
- possibly
- unfinished
- postecho
- detection */
+ while(j<ve->current-(ve->searchstep)){/* account for postecho
+ working back one window */
if(j>=testW)return(1);
if(ve->mark[j/ve->searchstep]){
if(j>centerW){
-#if 0
+ #if 0
if(j>ve->curmark){
float *marker=alloca(v->pcm_current*sizeof(*marker));
- int l;
+ int l,m;
memset(marker,0,sizeof(*marker)*v->pcm_current);
fprintf(stderr,"mark! seq=%d, cursor:%fs time:%fs\n",
seq,
_analysis_output_always("markL",seq,v->pcm[0],j,0,0,totalshift);
_analysis_output_always("markR",seq,v->pcm[1],j,0,0,totalshift);
+ for(m=0;m<VE_BANDS;m++){
+ char buf[80];
+ sprintf(buf,"delL%d",m);
+ for(l=0;l<last;l++)marker[l*ve->searchstep]=ve->filter[m].markers[l]*.1;
+ _analysis_output_always(buf,seq,marker,v->pcm_current,0,0,totalshift);
+ }
+
+ for(m=0;m<VE_BANDS;m++){
+ char buf[80];
+ sprintf(buf,"delR%d",m);
+ for(l=0;l<last;l++)marker[l*ve->searchstep]=ve->filter[m+VE_BANDS].markers[l]*.1;
+ _analysis_output_always(buf,seq,marker,v->pcm_current,0,0,totalshift);
+ }
+
+ for(l=0;l<last;l++)marker[l*ve->searchstep]=ve->stretchm[l]*.1;
+ _analysis_output_always("stretch",seq,marker,v->pcm_current,0,0,totalshift);
- for(l=0;l<last;l++)marker[l*ve->searchstep]=ve->filter[0].markers[l]*.01;
- _analysis_output_always("delL0",seq,marker,v->pcm_current,0,0,totalshift);
- for(l=0;l<last;l++)marker[l*ve->searchstep]=ve->filter[1].markers[l]*.01;
- _analysis_output_always("delL1",seq,marker,v->pcm_current,0,0,totalshift);
- for(l=0;l<last;l++)marker[l*ve->searchstep]=ve->filter[2].markers[l]*.01;
- _analysis_output_always("delL2",seq,marker,v->pcm_current,0,0,totalshift);
- for(l=0;l<last;l++)marker[l*ve->searchstep]=ve->filter[3].markers[l]*.01;
- _analysis_output_always("delL3",seq,marker,v->pcm_current,0,0,totalshift);
- for(l=0;l<last;l++)marker[l*ve->searchstep]=ve->filter[4].markers[l]*.01;
- _analysis_output_always("delR0",seq,marker,v->pcm_current,0,0,totalshift);
- for(l=0;l<last;l++)marker[l*ve->searchstep]=ve->filter[5].markers[l]*.01;
- _analysis_output_always("delR1",seq,marker,v->pcm_current,0,0,totalshift);
- for(l=0;l<last;l++)marker[l*ve->searchstep]=ve->filter[6].markers[l]*.01;
- _analysis_output_always("delR2",seq,marker,v->pcm_current,0,0,totalshift);
- for(l=0;l<last;l++)marker[l*ve->searchstep]=ve->filter[7].markers[l]*.01;
- _analysis_output_always("delR3",seq,marker,v->pcm_current,0,0,totalshift);
seq++;
}
-#endif
+ #endif
ve->curmark=j;
ve->cursor=j;
}
void _ve_envelope_shift(envelope_lookup *e,long shift){
- int smallsize=e->current/e->searchstep+VE_DIV/2; /* VE_DIV/2 is to
- match setting a
- mark on a region
- in
- envelope_search */
+ int smallsize=e->current/e->searchstep;
int smallshift=shift/e->searchstep;
int i;
memmove(e->mark,e->mark+smallshift,(smallsize-smallshift)*sizeof(*e->mark));
-#if 0
+ #if 0
for(i=0;i<VE_BANDS*e->ch;i++)
memmove(e->filter[i].markers,
e->filter[i].markers+smallshift,
(1024-smallshift)*sizeof(*(*e->filter).markers));
+ memmove(e->stretchm,
+ e->stretchm+smallshift,
+ (1024-smallshift)*sizeof(*e->stretchm));
totalshift+=shift;
-#endif
+#endif
e->current-=shift;
if(e->curmark>=0)
********************************************************************
function: key psychoacoustic settings for 44.1/48kHz
- last mod: $Id: psych_44.h,v 1.11 2002/03/24 21:24:01 xiphmont Exp $
+ last mod: $Id: psych_44.h,v 1.12 2002/03/29 07:10:40 xiphmont Exp $
********************************************************************/
static vorbis_info_psy_global _psy_global_44[5]={
{8, /* lines per eighth octave */
- /*{990.f,990.f,990.f,990.f}, {-990.f,-990.f,-990.f,-990.f}, -90.f,
- {0.f,0.f,0.f,0.f}, {-0.f,-0.f,-0.f,-0.f}, -90.f,*/
- {16.f,14.f,12.f,12.f}, {-990.f,-990.f,-990.f,-990.f}, -80.f,
+ {20.f,14.f,10.f,10.f,10.f,10.f},
+ {-60.f,-30.f,-40.f,-40.f,-40.f,-40.f}, 4,-75.f,
-6.f, 0,
},
{8, /* lines per eighth octave */
- /*{990.f,990.f,990.f,990.f}, {-990.f,-990.f,-990.f,-990.f}, -90.f,*/
- {14.f,12.f,11.f,11.f}, {-990.f,-990.f,-990.f,-990.f}, -85.f,
+ {16.f,12.f,8.f,8.f,8.f,8.f},
+ {-40.f,-30.f,-25.f,-25.f,-25.f,-25.f}, 4,-80.f,
-6.f, 0,
},
{8, /* lines per eighth octave */
- {14.f,12.f,10.f,10.f}, {-90.f,-90.f,-90.f,-90.f}, -85.f,
+ {14.f,10.f,6.f,6.f,6.f,6.f},
+ {-20.f,-20.f,-15.f,-15.f,-15.f,-15.f}, 4,-80.f,
-6.f, 0,
},
{8, /* lines per eighth octave */
- {12.f,10.f,8.f,8.f}, {-40.f,-40.f,-40.f,-40.f}, -90.f,
+ {14.f,8.f,6.f,6.f,6.f,6.f},
+ {-20.f,-15.f,-12.f,-12.f,-12.f,-12.f}, 2,-80.f,
-6.f, 0,
},
{8, /* lines per eighth octave */
- {12.f,10.f,8.f,8.f}, {-14.f,-14.f,-12.f,-12.f}, -90.f,
+ {14.f,10.f,6.f,6.f,6.f,6.f},
+ {-15.f,-15.f,-12.f,-12.f,-12.f,-12.f}, 2,-85.f,
-6.f, 0,
},
};