MR by all frames is not considered because the main
computational cost for MR is O.n3/, which means that
the data size n cannot be too large. We select a typical
frame feature instead of all frames feature, as the major
way in Manifold ranking. The experimental results
show that MF-typical is better than the baseline (which
means that MR effectively improves the retrieval and
ranking capability of traditional Euclidean similarity)
and MF-typical is better than MF-mean (which means
that the audio content represented by typical features
is more exact than the mean value). We compare
performances with/without RF.