From 317ffc203efc63333fc3b6a42fdb2887321a4325 Mon Sep 17 00:00:00 2001 From: Jean-Marc Valin Date: Tue, 9 Oct 2012 02:12:02 -0400 Subject: Tuning the speech/music probability model --- src/analysis.c | 7 +++---- 1 file changed, 3 insertions(+), 4 deletions(-) diff --git a/src/analysis.c b/src/analysis.c index 6742a85c..fde65493 100644 --- a/src/analysis.c +++ b/src/analysis.c @@ -409,10 +409,9 @@ void tonality_analysis(TonalityAnalysisState *tonal, AnalysisInfo *info, CELTEnc #ifndef FIXED_POINT mlp_process(&net, features, &frame_prob); - /* Adds a "probability dead zone", with a cap on certainty */ - frame_prob = .90*frame_prob*frame_prob*frame_prob; - frame_prob = .5*(frame_prob+1); + /* Curve fitting between the MLP probability and the actual probability */ + frame_prob = .01 + 1.21*frame_prob*frame_prob - .23*pow(frame_prob, 10); /*printf("%f\n", frame_prob);*/ { @@ -422,7 +421,7 @@ void tonality_analysis(TonalityAnalysisState *tonal, AnalysisInfo *info, CELTEnc /* One transition every 3 minutes */ tau = .00005; beta = .1; - max_certainty = 1.f/(10+1*tonal->last_transition); + max_certainty = .01+1.f/(20+.5*tonal->last_transition); p0 = (1-tonal->music_prob)*(1-tau) + tonal->music_prob *tau; p1 = tonal->music_prob *(1-tau) + (1-tonal->music_prob)*tau; p0 *= pow(1-frame_prob, beta); -- cgit v1.2.3