Hi All,
I expect this is a simple question - I’m still getting a grasp on how the process works and the language around it…
When I run a sound through fluid.bufnmf~ how are the channels ordered in the resultant multichannel buffer~ I’m assuming it may be that the first channel contributes most to the sound, then the others are descending order?
And is a stereo file treated as two separate mono processes?
Thanks in advance.
The order is arbitrary and if you rerun the process with the same settings you’ll notice the outputs Re similar but in a different order. If you supply a stereo buffer it will treat each channel as a mono input. Same goes for anything above two channels.
With regards to the first question about how the channels are “ordered”, it’s random. Rather, it starts from a random seed and starts reconstructing stuff, so there’s nothing to “sort” the results. Similarly, if you run an NMF on the same file over and over, the results will be in different places (and slightly different too).
What I’ve done in the past is then run some descriptor stuff on the results and order them either by loudness or centroid or whatever you think is important. @jamesbradbury whipped up a thing ages ago that then takes the weakest results and combine them.
As to the stereo/mono stuff, I think there’s a tab in the help file that explains that, though when I’ve done that I tend to just split things and manage it independently.
In the helpfile, there is an example that does exactly that indeed. There are also examples of pre-seeding significative bases, which allow you to guestimate where things will fall.
Thanks for the replies, that explains why I couldn’t work out the ordering!
So, my next task was to work out how to use some of the descriptor objects and how they connect up.
Which part of the helpfile was that? Just finding my way around everything…
the tab entitled “fixed bases” should help. There is an explanation and example on how a 10 bases over-separation of an extract is then ordered by centroid to pick one, recompose all other bases, and re-run the whole process on the whole file to get a predictible output.
I hope this helps!
Ok! Thanks for all the help.
After much tinkering and cobbling together of helpfiles I think I’ve got something close to working. The aim here is to split into 10 layers with fluid.bufnmf~ analyse for loudness with fluid.bufloudness~ and fluid.bufstats~ then to be able to sort through that data with dict.
If anyone has a chance to look at this it would be super helpful:
----------begin_max5_patcher----------
4587.3oc2cs0jaiar9Y6eEnzwUEumyLJ39k4Iu48TUpr4sMa4hiDGMzlhTGR
pY73Tw+1C.AoD0HPRHIHS6ncsDGJJ1n+5F8Mbg+q29lY2m+k3xYf6.+N3Mu4
e8127l5SYNwaZ962LaczWVjFUVeYyx1t993hY2X+pMQUKdLIa0GKhWTYuMHA
mOGdC.QnTymLn4cLdND7GM+p7sUowUUurI19SlM6Fvr6ixVMa2kjrrlZ42+o
aQD0dxUDsNtJt3iwYQ2mV+ygMemtgkjous0MSz9SZIV8YwlS9ue6aMucim76
Co45axP7KWvmi35WTX8KBxv7LrgsEzKk4ovyl4eHuXcTcqj+8DNXLGvgjTCG
nKFNP+PAGYwOqaVsMop3uTe+mkj8P92.2u8gr0O7wGiS27wxpnpR1oiavZbC
QIiCbZQil8znWZRY8m6NgOGX9WRVUOnNdl+8xTmCNNpUElhFXqJrenspvfzq
nUEB+mLqJTsAjqmUEh3mSqJODGUssHdXCKtgtFCKbzjZXgHuxFVteaUUd1Pv
CAgNFdv35XXvz126GcFPqhFHKLnqhEUJFGXKp3ensnRH3qoE0e1hSyodevrn
hU+bZQUaL8krpGYmNx0XPkIlTCpXwU1f5nJVXk0rBVhCiYEL6GJUoE4qWGqg
+WqK82Rid49nEeFnIJHBTpalw4fGRzMy9AKhDZsAiEVsGKZA6fVG.EjSP5BC
W+DMHVj7kuAPP.Z.tgiHy4c5X.kDStK1XLHP4b1XZBkIqxhR200vMHfbCBvf
Ihi+5xnEeyCcbt0uAk09t6FLzYCFGLoVZxSwyWEkjsqM+TTQltqROe6.bDCY
SyBaslgU9Is1ezAlp5WHh14ePyE5VSTVVb5h7sYUc6T6pCOxerj0b1xnmhW9
wnJsRrNdv38GU1.pMnpA2R2Fm+P6oaOe2Vh192p9w1Ctz0ZoTswm4PmeaRl4
auU.c98kOlWT4IkZEJttOayRpJqdwhdT62Wqf09w2UaDsJZg1p.RdssJXYvE
UEoCEb.BNmIIDjp4ES1w7N214ZvNUswDbHqVtH5P+cKxS2tN6Pd7.737SosH
+4.GbdOpLZ4527RYoA+ndDVZithacD9rqbRWt4yUE44OE6NPSsydvGRyy23G
PnpABNGOmr+EVBI92G5+JNxs7k4T9RbqGiNv6f9FnQ6AbMbeTYxhJswTSbpZ
CozVas5nLSSyedUZ98QoUwq2j20RrMFzrpE4EFwYRd1AeqQxGmszJjpC9a15
xcbm86KqhJp58JVmuz5gntA15bXVdQxpDMRkFmsp5wt+5pjEet6Mn8J20zQl
3O69BQau1MOFUFOzMaShV2sGd8+eaTZR0KNZrUIqiKqJh0+1laN3ONeGT56U
Yzp3i5FdqBNWxwBHBbqjNWxDD8A5.cjZC15C4x4HARnTfa0YwIEbJWeHUMmh
gLr9GQEZa6HtTej1Pt.SELvsD4bASJjlCEywDNWtyzaZRV7tfZPvAbaHsNHz
QLabGJD6p5+.YM4tO.8DB1DEPzUAXZwFGH.D.BHAT.dHCZhVFtYvB8ohCtYX
7Ulg6wn9WSmWpCPaHdzlW.R2CxWuW0QP6lMOkjHBnO5seMAffCxkVWRLtuLY
cp92r+ydKlATcB7LInhVsK6saFiwqEuJ8GFQLj2XrT6XAnMkbAxZHeZj0FU5
zjEwCGJhkskJUct9RQiSBL8F.icQrMaZX6MEwFevfUwC0aFgqkoRcmYiEZx4
ZuBhmMwU9GyU08XUbCmbwEnSMkSOidDoKxWeu12KnovBk2cGHAb2cKiKRdJp
RmJs9Lv6tacbTF3C5H9VsJtnbP09FwufZGk9CT6IvKwztRbJQsFPO16AGDrG
rYezMNCoA2OdIgcyPQmPLsGaFHcTS7tujbMhxN4NVJ3zDGvRcVAfHc7yuTlT
5g8CF8XL.Sviq.YHjNj5nhWZqyVupTR4IfEzSKOnXs0kkGDW+.c6OndWQa1r
6qO+n66q96OEWnSFIJE70778iOgNMa8EGY.tNxBAtoNLl.+Q5bmknip6tygL
0JmHBwwxPG2BSaaQdZdQaFSZys6eC4tbwCO.NcYnuKk9uBLX4tHbakYLILow
RQak+oJKVbd9Hmn5zTAt27eCNn4Bksrk7i4WJ8zi8sWupDzIfAj.1Op5wXSg
pdHt.7a+cPRIvbhx7sEK5b9+7i4aFoOF1Zui1zIC2VPTgwbGzqtaJV8sfvTG
6y0083r5uQvSTGt9bKifyGAZQbXePK+39bNQVRcreDkc.6X7V0XsuWA6L61R
FtzC9gjn.pJaiDFj+.3AiSnx6.u+8MZx+O1y.9+.OmjsL+4xjuF+Kf+LPqW+
a0Gd69nAOHjGhWZ8cbsPgLuT1k1bFoXliXj309cLYZeYJ6T0OVdWVFqsmjl+
rOtXbAL1hNbN5phowESZdzRSGeS4S6NGY5ksILGIbnN6tnpqLa2mQsmLQkAd
GxGAsNapiDz7yMea50NsfWMgX5yh8qrJv7yNcqRfq.Nu3D2oj.NwZ7yzQHSe
EiUd5n7zwcg85n5zeQmEt6Lxd0D4erGaQweYSA3cOf.+u52wZme0Updvk1iM
IGhxF8KkzN.Qj5AocNlcV.ibhRc+LmP0snfqvRm3ITsDckm9e80YrLtpI4.m
v4Mf6S1jmFULbM1X1IbMQvcUiD5Y1uSRlF0KKfbnB18QkincQ3LGShNjM4RD
+TVHP8mb4rqaUYGcV2SURKSRZRcgqLi2Isdt.cIS5dIdBmy8mGaSvXGo2dQq
7.3DBBlxTN+oj3ghomgalA0cwAE1Uo0sCLLQhmq57RJ4tWnZW44G6YJgU7.K
g47obok31b2CoaSVNWarqdsN9MvG1WznicHn+VyU4XMRB9ftEVOXDCOfLtQZ
WYKQUP+KL2d2rNwc3Utpb84esHdSZzh3S2wAzNXeT44l+j3J67zLRCaSFjwr
xOjcZTfNs3uM29Z574miSV8XkY5acLWhwcQ7O9bxR6blpdZ43emo1BFku3yW
O3fQsISZWhZmKb7X9FSMubgEjedvhVUiKCKzTIdSbzmcAFre9TLtr9I6qGpC
3X+pbapgiOUNhlQSwjT34R19Ymi1pnNCdgclOnSfi0cZNhn9fT2mWncO00Oa
WykAywbakVRRi2sH.14h0De07OUNKfEK23.D77iwYfk4YwCuzzapFEuwyKBo
0yDNhwQoByXzfgScbcsC7GVPCbnbXzOvgxklucYlNjjuY1VWnPmwzsalr+g1
n65Kju8lVZucMdg.T5fS0NDRzD+BZNG9pdyBHJTg2gHW4v6NgJEL9NFhfRli
oJknseHkVOiZY0y7FTaw4jlTq8YfRFutAbzrIY5b4Be7Yk5pPJGHDBYWbaBX
HvD1OPXhekuzoZyNPgFBPAK+grBSb5EuqNvlPq08451LxAkfpbfckhLTcWnN
xYlamWqCsxiOokU943.1vKGJSeLwNf6ylsaLmx1ubH+0hjnTveIOc4tQoOdU
zhWNH7BzoV8.dKF7Jjv8j4700guHea1R6biiuaIcNKYUVtltoIMw51WbPDpB
huwwA8MuU7Zz6teUGhT6mDNWHkXRcDWTAjIkGRBmB38SNkYShuvhELv+nHd8
8oQub6uF8a29qKx2VVkr32zIEst71+57mid5LppNBdruxlsDjK0DHcZbKrKz
MM34tFb0f4Gz1R1niyOqpzrhc+P83R3XjJ.enwOaeg78PUoFXzns49fo.Liq
66Pkmbw+sYsJQrPELGdZhkaeYP2n43hnzxGi1D+sKLz4fgyNCgl6y9KmentZ
hJP5fSsDEAamA.cWU1NmJ+X07yaLr4zoYPFsyY0EowQEmppfNoqiA.Nmb5yh
UK4cBKrooW3FyNwBDLTRkLsw+ivDlq76QR7YV8bt3GmgddzDJEH9wiJGi3ZV
NvBRVBb4z3ibUbkobmfxn0aFZIavEDGZHPG.hj34B.p2AVRMM8S190gphJGq
b..tzHnvvstO4Syp97L1zVcZV0oBBhpNX.rUB7zMKgDvIZVB0LJlf2Q.uCAd
mFBr5Bm1DEP5zjD4bmgnL0zXi9S4IYmLqq8Yc7bjfdtyBZAZZXcGKOyEud4Y
B5txLIC.TRp7XfRHc.TXA4BVmlB2ITPu1dqLQ3kpikI8TikgCINm8pLeMT2Z
CoODgcJoXQCo5y1rpAWhyt0IbEvuOyWyVLXLyptG8.10FN1lYi1c9w++oZew
8xaFeJQ6cFG3DLOk5mHB55cbvU.naXCwOWyxCVBI1zDe7WSAI59W.wohCboq
PCUWfQWJcZ1SH9TYiuGv60WA3uTmO06+SGGc3e5WluIN9yuO52g+w+7edCT+
OPzui9ie4WNYvii6c9OdN91YSyJbx20Cu63ab411mgj5DVb7B9+8t33GPlLu
1IgYWc3NPY0xkwOo+7ywOaFd+6.edaQUtVfcGHM+46.qSVdG3wjUOd2IaJzk
DzKOHcDgNEatKkh3puz0Fs1ZMyeaOBT+bW1dHL+LCTu9tUuBVe0S4mZFwb9C
QDaQpaZK61E6A6avKiKqRx1sHk98NOgP5bU6mhTNEDAl3lmh.tHN5xH9q3o9
3bj8pB.s3iRKXvnkbTPUEBZUKZFmVrfQqQwPyi9j.QqwULEAgVd0I.Kb2I.e
YDu91Nt.DGBFslViK.gAiViCpAoS.l6AsDAQuDy7gsPggVHunkABQWJsP9It
PWEuPl8UV.R4E0UWLmVSLoWDSFFhI7hXhvPLtWDiGFhw7hXrvPLpWDiFFhQ7
hXjvPLrWDCGFhg7hXnvPLuMoD.hMpADydXefLfH8hVgw9gvKZEFyGbunUXrd
v7hVgw3A0KZEFaGDunUXLcf8hVAxxgWzJLFNfdQqK2tA2KZEjzmZCTaD6gvf
fgJuBmKHQ5ijdGlJ76SXpnqR8Jn9PaZ.3ReJKhLDID5SGsfj3IB5kXiFDUen
WlpvgfVP+LeDjtYPuLACCRlmlMF9QcsnBQOZlW0.DFLRMF.JXghRiAelQyty
EsNY4l7jrplxHyntFAKliIbfT3Zg3b7UFBdZrHa3GVIsWySLjqkLD938hHE0
9724f4Sriq7B4ItOVzYgvBnT5UmW1Uo7JPubnDh3oTd4dFFBDU4SDN0sGz2C
Ym5PuNAJjCl56UgM49jLgfOT2acXxGNPcRFSadh6XFuvHNlNUNtxKjoDd4dN
DgrQPd.exPTpCoOEElPcqMBubROVABjgH+EuXRI7pLpj9nyPCQWNuJuOIH5L
9D9sHDii.wKdREpdbiYU1ryDd4TxGivzPPo1axvTJDYSH8gRAYTrn9n6Ejww
kH9dwSRuRZIDZDBe3IQHrQH7oX.zP3qT3yvsKDghRHenDJDTB6CkvgfRDenD
IDTh5CkngfRLenDKDTh6Cktzg3P3iEVwgIU+p.pkLwwQIKDNd9FgzuN3RqC8
1wkdoLE1WyQWbGJeBjfEh.ILKVqQGNufjhkO0+mKCEkFC73gvtG2mnk4rPkf
5npdvPQoQkSpgREVRfGW8NpzwCXCtj30Udo7D0WqQWrFgOw6wCR+IOb49pJ7
8JwDm6PLw3tdVPf86Juzfk8RIODAgQ8QIWDjjZ7Q0iFBSDD3IDVdO0vhwX06
ShX6t0EiZeR1c3ehg1Gtxlmpzrl+7RKF.6xa6ZcycMm.25HACYuFsNbvvtVA
Mo4yC+yqPaG4Cxdkl1kdITuNEYyKmDG6L2tZKh1r4o3hxlqtlHyVG8I6tvk7
l5+LIqyVT1rh3mRZudj8TQEKdLoJdgYC.pdcf7klcXlYqy07X11jF1TyeZZV
V8R5qWhG61rzVG8EtYsuTyNlm4VQaSqNDB1scj8wjLSiOtcaIqdGCq8st6SX
Guyhs+M378YKEsXQb1ga0YRyUHIJLhWuejAURHs4taQyaZDWElMnJCq0Ymfa
1ZcqOoU+pq38.FtdUxLDKePShhILrpdWWSfYz5ivBJCE5l0AqxnvIKdHIMcG
C0cEo0tzhlspHZYx9cOvF9GsSlXhTPbiyi5PpleFt8mseuj63iN9m09qvJHR
UK6IMxdyQ5Sgd0uJJakEjwhc6ndlmTR4axKZ6Np+opcW+1p7cL5QKkN2huC6
c3mTztZq7V6pVlQYsareTMOSTmslk0JS8BWyznJ2DYMdUu91d6+9s+GfyInX
0
-----------end_max5_patcher-----------
A couple of things I’m unsure of:
a) If I run a stereo file through it I’m getting a 20 channel resynth buffer as expected, but then 40 channel buffers for features and stats - why is this? (I’m assuming I need to read up more on what exactly these are, but pointers appreciated)…
b) What do I need to extract from the dict to measure loudness for each channel? I’m guessing (admittedly with no real knowledge of this sort of analysis) that taking one of the mean values will get me there?
I’m not sure what the relevance of the derivatives 0 and 1 are:
If you imagine you have a time series that goes:
0 1 2 3 4
The derivative of this is:
1 1 1 1
which describes the change between consecutive samples.
.bufstats~ will calculate the statistics of both of these time series and give you the results in separate channels.
The idea is that you can capture the variation within a time series by analysing the derivative time series too. There are pitfalls to that assumption, for example, imagine we have our original time series and a new one, 4 3 2 1 0 to compare. They have the exact same first derivative time series and we cannot differentiate them unless we start looking at some other metric or form of analysis. Further to that, adding more derivatives doesn’t help, as the next derivative is 0 0 0.
A new channel for each derivative of analysis is produced in the output 
I think your patching overall is pretty solid - can’t see anything that would hinder you from getting toward your goal of analysing the components of .bufnmf~ by loudness. Its likely you don’t need to enter a dictionary at all though and you could use several zl objects to sort a list of indexes that describe the order of components.
Something like this:
----------begin_max5_patcher----------
618.3ocqUksaiBCE8Y3qvxOShvl89qLppxg3xPEXiLlNoSU+2G6qMIgrLEUk
HEi456x4bWLeFFf2IOvGwnmP+BED7YXP.HxJHv+d.tmcntiMBpg64iirFNNx
cllePCxSPwHBJCQQoym00J30xIAn.0KTL02J535wyE1tG7gb2aaHE3SZJmzy
pR7RchzeLvcnFiQO6OZfoq+cqn4EEuV6NMiltMNBQhyKrOyhsqIYaiQOas4q
vP6RzJ4tf+GCDuh5+sCMJUZ7ZYH81LjdWFF8crLIwwRRl8YJwtRo+PVduJ7l
7XzlLyeZJZCIM0Tv2jjuZVSdn0UGiqJxAhSytIiAO.sgWzhCdzJeYZXTNopm
i+LnMzbFH64i5VAS2JEmqD0ozMS0qNRVmP9tHU3T5JJxFFdmqF8ZCAwTSeSp
ruVFAu1JbuBIbrh+d6r9NEXJSRVaxvSJ.X3C4tIYbubOWIlZAqCszyDxQ8Gc
WlVErdmol9obaSFvFCYdkM0oWlArsU0xNo5kVgE69jALfdZgr8T1mUWyENal
0szpVYREkja2UDWUFmtzpcMmYAw6S+R71r4jYjuZoLg.nFvCOA5MnGXe7EU2
EDFZ++eTdAzSoIYzJKFRJnYovNZQZFwC+GFrVLN+npE6Zdssq6HgBN5Kiy7i
v3FEaeqA294dO+IGqc4leEQ2b2YgxaFc1rhxRSlK5l6t1rYqnUwjJnGIw2iX
2YDQtvJlnwkjoEVR6kNnjCla38yKFSqNp+jVdjn9Rge37tkukSGqqJ5tdb0c
WPMK0zVESgcFNav7OsyxcIC7EBKnFGXt6tfOjD9U3+.Sm1cW
-----------end_max5_patcher-----------
Thanks for explaining all this and the feedback on the patch- starting to see how these all come together…
Going to spend some time running some sounds through and listening to the ordering.