Freesound Forums

When I create Samples I work hard on it and before Sharing them,
I check the Frequency Spectral Graphic of the sample to see
what Tags I can attribute to them...

But when I search for samples It's Hard in some case to see
what the sound is with just the Amplitude graphic,
So it's useful to see the FFT Spectral of the sound,
because when you have more that 100 samples to check when
you search you will surely not Hear/Download all them...

So Bram, my first concept I propose is to Add the FFT Spectral Preview
for samples, for instance:
http://img245.imageshack.us/img245/6581/z2sc6.jpg

But this reduce the view of the previews so I think of an second
concept that let the possibility to choose between Amplitude mode
and FFT Spectral mode in the left frame menu, for instance:
http://img131.imageshack.us/img131/4245/z4wm5.jpg

So when you are in Amplitude mode:

http://img440.imageshack.us/img440/6855/z1py7.jpg

And when you are in Spectral mode:

http://img131.imageshack.us/img131/5767/z3ze8.jpg

Spectral can be useful too to determinate the real quality of sound,
For instance an 16Bit-Depth 44100Hz sample:
http://img245.imageshack.us/img245/2526/z5tr0.jpg
And resample to 22050Hz and after ressampled to 44100Hz:
http://img440.imageshack.us/img440/595/z6xi6.jpg

So there is more example like that, when Spectral can be usefull

Spectral can be useful to see for instance when you search for an
Drone/Synth not just the must dominant Frequency but too for
the harmonic frequencies.

I really would like hear comments about that :mrgreen: ...

I also use spectral views often, so I would also love this feature. You can tell bird species from spectral views, amazing hock:

On the otrher hand, don't you think that would impose a severe load on the server? At least my computer nearly freezes up when it paints a 5 min sample at 24 bit. I know, I shouldn't use 1024 band resolution, but it looks sooo nice ! More specifically, you were thinking of generating spectral views, caching the images, and serving them along with sounds, don't you?

I'm not quite sure what you're talking about, dobroide. I thought that the waveform pictures where captured via some porgrim the mods use, then the image is put along with the sample and the picture stays stationary. That is the pictures are .png's but don't have a frame rate so I asume it wouldn't take more for a computer than simply loading the image. :?

When I first came on freesound, I thought it would be more convenient to have some sort of player that replays the image when you play the sound, especially for long samples where you can't tell what you're listening to. Now that would put a load on the server.

This seems like a good idea to me.

On the otrher hand, don't you think that would impose a severe load on the server? At least my computer nearly freezes up when it paints a 5 min sample at 24 bit.

Like hello_flowers said that"s Picture(*.PNG) files pre-calculed just one time when they are shared on server

But I think that calculate the FFT Spectral with more that 512 or 1024 Dots is just a lost of time
for samples with frequency range <=22050Hz (Sampling rate<=44100Hz)...

This summer I will make an VLF receiver
and I hope to be able to record some Strange natural VLF interferences.

These interference are hard to see on an amplitude preview:
http://img440.imageshack.us/img440/8597/z7lf7.jpg
But with an Spectral preview they are more easy to see:
http://img245.imageshack.us/img245/3248/z8ab4.jpg

And I also working on MatLab to make FFT Spectral Picture algorithm
and make new strange sounds :mrgreen:

how beautiful the last one, HardPCM, makes me think of whirls of sand from the desert... Spectrals are often charming on their own, like watercolors or abstract paintings.
Have a look at this one:

http://upload7.postimage.org/707226/Clipboard.jpg

Cool!!!
Seem to be an bird, what is this bird?

Or just do some funny thing with the noise :lol: :
http://img338.imageshack.us/img338/913/z9yb1.jpg

But Spectral can be useful to detect if there is an 16Khz TV tone interference in the recording,
and Detect if the recording is from an TV or VCR:
http://img369.imageshack.us/img369/5580/z1bo6.jpg

Now I stop to post big JPEG and wait for comments :evil: ...

yes, nice feature ...very useful idea :wink: ...cya

tro

HardPCM
Cool!!!
Seem to be an bird, what is this bird?

Actually several birds: a Hoopoe made the groups of 4 tiny slashes arounf 500 kHz; the vertical rythmic bands are from a Blackbird; intricate lines that remind signatures are from Starling, etcetera...
I'll post the correponding sound sample later today and link to the image.

Now I hope Comments from Moderators/Bram :twisted: ...

Would certainly be cool, but would also constitute a major feature and take a significant amount of work.

Doesn't the "spectral centroid" coloring of the waveform represent part of this data?

halleck is correct, this would mean quite a bit of work, and at least 48 hours of server processing (which would right now mean the server would be... dead all those hours).... So, not a hot idea for right now.

The coloring of the waveforms uses the spectral centroid for colors, which IS related to the "pitch" of the sound somehow... ( do a google for spectral centroid )

But I agree it looks kind of cool!

- bram

Halleck
Doesn't the "spectral centroid" coloring of the waveform represent part of this data?

Yes, the color of the waveform is just the higher frequency in the spectrum (not sure of that)
in function of the time of the sample

And I know that there is more that 30000 files to scan,
So if for instance you scan 500 samples per days,
30000samples / 500samples per days = 60days
and the new samples are scanned when moderated...

And I know that there is more than 30000 Small and Big preview pictures of samples,
so that"s double(2x) the count of pictures (60000*2=120000pictures)
and the size of they take(2x) :evil: ...

So that that need some works,
but Freesounds is now more serious,
and with more that 30000samples,
I think that we need more serious tools for preview/searching samples...

( You do realize that I'm the guy who programmed freesound, right? )

The trouble is that this machine is hosting 15000 people/day and as such it's CPU and memory are at the maximum of their potential almost 99% of the time. Doing something like additional spectral processing would slow the machine to a crawl.

If you would like to see the effect of thousands of people downloading all the time, have a look at:
http://iua-share.upf.edu/ganglia/?c=IUA&h=iua-freesound.upf.es&m=&r=hour&s=descending&hc=4

- bram

Very cool, and you really running that on one standard PC?

My suggestion is to do the batch processing When:
SpectralProcessing% + OtherThings% < 75% of CPU power
do some test to find how the algorithm take to the cpu (SpectralProcessing%)...

that"s take more time to process all,
because I see on the graphic that 30%/40% of the time
the cpu is between 50% and 75%,
so that can take around 200days,
but we are sure that the cpu will not overHeating/BUrning :mrgreen:

That's you the master, but I think you are able to coding that

What is this?
http://img485.imageshack.us/img485/3575/z0mm6.jpg

The HD will be nearly saturate?

31209samples / 480.22GB = 64,9889samples per GB
64,9889samples per GB * 499GB = 32429samples!!!

This is an average estimation but that"s near :evil: ...

Some form of distributed computing (Seti-like) could do the trick. Anyway, I doubt the coolness of having spectral views justifies the effort to implement this...

I estimate that the additional preview pictures (small and big)
will just take 1.6GB (i'm not sure)...

But now the problem is to process them
without make crash/Burning the Server CPU :?

And how would you search a sample
when they are more than 1000 for one tag/word?

Or just have the client run the computations instead of the server before uploading. But this would require developing an actual client.

Very easy for us to sit and talk about it but in the end it comes down to a crapload of work for bram.

IDEA:

Why not cut off the new server for a few days to configure all of the images on there, then when the servers synch all the new samples [from the "old server"] are automatically put back into moderation, that way, when they pass moderation they are automaticaly assigned amplitude and spectral images via some script on the new server to say all samples must have both images to represent them.

I think that might have ended up a bit confusing, maybe's...