Hi,
Thanks for the correlation analysis. In my opinion, it is normal to find such
high correlations between particular features: the descriptions of spectral
distribution (2—7) have some redundancies. Subband flux describes different
bands separately, so it is normal to find correlation between audio bands.
Best,
Olivier
21. mai 2018 kl. 23:22 skrev Enzo Tagliazucchi <nztglzcch@xxxxxxxxx>:
Hi! thanks for the clarification.
The correlation is not high between all features - a few of them are
uncorrelated with the others, but I believe that the correlation is overall
high. I am attaching the correlation matrix between all features, and below
is a list of the features we computed:
1 zero crossing rate.
2 spectral centroid.
3 high energy-low energy ratio.
4 spectral spread.
5 spectral roll off.
6spectral entropy
7 wiener entropy
8 roughness.
9 RMS. loudness
10 to 20 are subband flux
21 Pulse Clarity
22 Mode clarity
23 key clarity
On Sun, May 20, 2018 at 6:00 AM, Olivier Lartillot <olartillot@xxxxxxxxx
<mailto:olartillot@xxxxxxxxx>> wrote:
Hi Enzo,
f = mirfluctuation by default gives fluctuation strength decomposed into
channels (critical bands).
mirentropy(f) gives the entropy for each channel separately, so you get a
vector.
f = mirfluctuation(…, ‘Summary’) sums along channels. mirentropy(f) is now a
scalar.
About your last question, what features have you computed?
Best,
Olivier
18. mai 2018 kl. 17:25 skrev Enzo Tagliazucchi <nztglzcch@xxxxxxxxx
<mailto:nztglzcch@xxxxxxxxx>>:
hi Olivier!
thank you very much about the clarification on the concept of fluctuation
pattern, we've read the citation you mentioned and it is clear.
w.r.t. to fluctuation entropy we meant applying the mirentropy to the
results of the mirfluctuation output and the result of this is a vector
instead of a scalar, but we will check our code once more.
One brief question: we have two records for which most of the features show
a high linear correlation (R>0.8). I am wondering if this lack of
statistical independence between the features is normal, or it would be due
to lack of proper parameter tuning (i.e. window length for the
window-related features), etc. What is your opinion?
thank you once more
Enzo
On Sat, May 12, 2018 at 11:56 AM, Olivier Lartillot <olartillot@xxxxxxxxx
<mailto:olartillot@xxxxxxxxx>> wrote:
Hi Enzo,
Sorry for the late reply.
mirfluctuation is not directly related to the concept of fluctuation
strength (which is an indication of roughness).
mirfluctuation is related to the concept of Fluctuation Pattern (also called
Rhythm Pattern) (cf Pampalk, 2004), which shows the rhythmical periodicities
in the signal. More information in the Users’ Manual.
It would be good to have an implementation of fluctuation strength as well.
What do you mean by fluctuation entropy?
All the best,
Olivier
4. apr. 2018 kl. 19:15 skrev Enzo Tagliazucchi <nztglzcch@xxxxxxxxx
<mailto:nztglzcch@xxxxxxxxx>>:
Hi all,
we have been working with both functions, that are expected to return a
scalar value containing the fluctuation strength and entropy.
However, when we run the functions with the following parameters:
mirfluctuation(A,'Frame',3,'s',67,'%')
we receive as an output a vector of values. Can you please tell us what
these values represent, and how can we operate on them to find the
fluctuation strength/entropy?
Thank you!
<correlation_features.png>
