eden-emu/opus
15
0
Fork 0
mirror of https://github.com/xiph/opus.git synced 2025-05-31 15:47:43 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 4

Makes the speech/music probability estimation mode conservative

Browse source 
This is done using an adaptive beta and an estimate of the speech
and music detection confidence
This commit is contained in:
Jean-Marc Valin 2013-02-28 15:30:51 -05:00
parent 742aac1056
commit 73142b100a
2 changed files with 37 additions and 6 deletions
Download patch file Download diff file Expand all files Collapse all files

39
src/analysis.c
View file

@ -180,6 +180,7 @@ void tonality_get_info(TonalityAnalysisState *tonal, AnalysisInfo *info_out, int
psum += tonal->pmusic[i];
for (;i<DETECT_SIZE;i++)
psum += tonal->pspeech[i];
psum = psum*tonal->music_confidence + (1-psum)*tonal->speech_confidence;
/*printf("%f %f\n", psum, info_out->music_prob);*/
info_out->music_prob = psum;
@ -479,19 +480,22 @@ void tonality_analysis(TonalityAnalysisState *tonal, AnalysisInfo *info_out, con
{
float tau, beta;
float p0, p1;
float max_certainty;
/* One transition every 3 minutes */
tau = .00005f*frame_probs[1];
beta = .05f;
max_certainty = .01f+1.f/(20.f+.5f*tonal->last_transition);
max_certainty = 0;
if (1) {
/* Adapt beta based on how "unexpected" the new prob is */
float p, q;
p = MAX16(.05f,MIN16(.95f,frame_probs[0]));
q = MAX16(.05f,MIN16(.95f,tonal->music_prob));
beta = .01+.05*ABS16(p-q)/(p*(1-q)+q*(1-p));
}
p0 = (1-tonal->music_prob)*(1-tau) + tonal->music_prob *tau;
p1 = tonal->music_prob *(1-tau) + (1-tonal->music_prob)*tau;
p0 *= (float)pow(1-frame_probs[0], beta);
p1 *= (float)pow(frame_probs[0], beta);
tonal->music_prob = MAX16(max_certainty, MIN16(1-max_certainty, p1/(p0+p1)));
tonal->music_prob = p1/(p0+p1);
info->music_prob = tonal->music_prob;
info->music_prob = frame_probs[0];
float psum=1e-20;
float speech0 = (float)pow(1-frame_probs[0], beta);
@ -526,7 +530,30 @@ void tonality_analysis(TonalityAnalysisState *tonal, AnalysisInfo *info_out, con
for (i=1;i<DETECT_SIZE;i++)
psum += tonal->pspeech[i];
/*printf("%f\n", psum);*/
/* Estimate our confidence in the speech/music decisions */
if (frame_probs[1]>.75)
{
if (tonal->music_prob>.9)
{
float adapt;
adapt = 1.f/(++tonal->music_confidence_count);
tonal->music_confidence_count = IMIN(tonal->music_confidence_count, 500);
tonal->music_confidence += adapt*MAX16(-.2f,frame_probs[0]-tonal->music_confidence);
}
if (tonal->music_prob<.1)
{
float adapt;
adapt = 1.f/(++tonal->speech_confidence_count);
tonal->speech_confidence_count = IMIN(tonal->speech_confidence_count, 500);
tonal->speech_confidence += adapt*MIN16(.2f,frame_probs[0]-tonal->speech_confidence);
}
} else {
if (tonal->music_confidence_count==0)
tonal->music_confidence = .9;
if (tonal->speech_confidence_count==0)
tonal->speech_confidence = .1;
}
psum = MAX16(tonal->speech_confidence, MIN16(tonal->music_confidence, psum));
}
if (tonal->last_music != (tonal->music_prob>.5f))
tonal->last_transition=0;

4
src/analysis.h
View file

@ -65,6 +65,10 @@ typedef struct {
int analysis_offset;
float pspeech[DETECT_SIZE];
float pmusic[DETECT_SIZE];
float speech_confidence;
float music_confidence;
int speech_confidence_count;
int music_confidence_count;
int write_pos;
int read_pos;
int read_subframe;