@37:17 at the Hann function you need the minus one at the end `float t = (float ) i / (N - 1);` because if you avoid it, the last value in the sample will not be 0 but a fraction, which is why you can see the seam between the first and second copy of the sample at @41:47
When you see lines above those are usually called overtones/ harmonics. Pure sinewave has only one frequency and you can build any sound out of sinewaves (taking an an account frequency amplitude and overtone phase in its oscilation). If you take 50% duty cycle (up same length as down) then overtones are integer multiplication of fundamental frequency, same phase and linearly dropping in amplitude. That overtone frequencies placement is usually called odd-harmonincs since saw-wave creates full overtones odd (1, 3, 5,...) and even (in-middle of odd harmonics 2, 4.....). Sound is a fascinating phenomenon. You can observe it in different domains and find out new things about it.
You could try it by generating pure tones in audacity. If you do so, you could generate sines waves, square, saw and so on. Your visualizer should show the "base" frequency and all its harmonics. It would be a nice "test" input file that you could even commit.
@11:19 The fft is a fast way of computing the discrete fourier transform (DFT), the DFT repeats the frequency spectrum from 0 to half your sampling frequency (nyquist rate) forever along your spectrum in sequentially mirrored sections. Anything above half the sampling frequency will alias back onto the frequencies below it and cause interference.
I was really confused with the lack of aliasing in the output, until he removed the 20.0f, when the signal mirrored I was relieved my control theory wasn't as rusty as I thought it was lmao
For an amplitude don't use natural logarithm, but base 10 (log10f from math.h). Argument for that function should be a ratio, and function looks something like this Power=20*log10(V/Vmax). When V==Vmax you have a ratio 1 and calculated P is 0dB. When V/Vmax equals half (0.5) calculated power is -6dB. You can draw an amplitude from -100dB up to 0db linearly on vertical scale
But log_10(x) is just (1/ln(10)) ln(x) ? I guess if you want the heights to be interpretable as decibels? But, this should just be rescaling everything by a constant, right? So...?
@@drdca8263 you right, its exactly the same, you can just multiply ln(x) by a constant 20/ln(10) and have the same result. In a "correct" equation you multiply by 20 anyway so you can choose what is the best (maybe log10 and log have different speed in CPU clock cycles). Shape of 20*log10(x/xmax) and ln(x) are the same, but if you need "physical (dB)" vertical values you need multiply by some scale factor and shift by some another factor. Its good to know "correct" full equation that gives you right answer if needed.
50:15 log(a*a) = log(a^2) = 2 log(a) and since you're rescaling the bars based on maximum amplitude this does absolutely nothing apart from maybe avoiding the square root
I imagine that the reason for it being N-1 is because the largest value you put in for n is N-1, and you want the last thing to be 2pi instead of slightly less than 2pi?
The frequency of notes aren't linear, they double for each octave! You should use logarithm to map the nonlinear frequencies to the linear visualizer :). I think they did that too in the website's visualizer.
Maybe I misunderstand the way the code works, but I wonder if some of the weird artifacts occur because there is one thread writing into the input buffer amd another processing it. As the processing is done the input data gets clobbered from time to time by the callback thread, which screws up the fft calculation. Maybe?
@@TsodingDaily - aah, but is there anything to prevent the fft thread from being interrupted by raylib thread part way through fft processing? The fft thread doesn't acquire that mutex (does it?) so the raylib thread will just acquire the mutex and do what it does. Like I say, maybe I misunderstand and am wasting your time.
@@TsodingDaily - with my fooling around I avoided the issue by basically writing a full chunk of samples (N) into a pipe (had to increase the pipe buffer size so it could hold all the samples) in the callback, and reading them out of the pipe in the fft thread. Maybe not that efficient, but whatever - it's just fooling around 😀. You can probably imagine a much better way than what I dreamed up!
log(a^2) = 2*log(a). Since you normalize it, it should have no effect. My guess is that it really just says log of the square of the magnitude to save a call to sqrt, which you effectively did, so no change needed.
I knew it sucked but didn't want to spoil your enthusiasm. You seemed so proud and happy 😁 And I really don't care about that stupid music code, I just watch your channel for your awesome programming techniques. Toll!
Education mainly I think. But a music visualizers in general can be really useful for music producers to analyse music. However sound is just 1 kind of wave. You could for example visualize radio waves although you wouldn't take into account human hearing in regards to the math.
@@Phoenix4_Trade Same for me, he's awesome. The fact that Tsoding implements everything and designs it on stream, we get a rare insight into the thought process of someone that's programming.
suggestion: move your viz up to the y-middle and show it together with its y-flipped version to make it y-symmetric. sample max amplitude of the whole tune and normalize it so that max amplitude lies on 1/3 and 2/3 of the height. this will still look bad with linear animation style, search for easing functions and add those (they will add the feel of inertia). also fix color, it is horrid. now you will have a cool looking and naturally moving musializer.
I'm really loving these shorter Tsoding sessions, like fixing or improving code developed on the long twitch stream session ❤
@37:17 at the Hann function you need the minus one at the end `float t = (float ) i / (N - 1);` because if you avoid it, the last value in the sample will not be 0 but a fraction, which is why you can see the seam between the first and second copy of the sample at @41:47
I love tsoding. But I was mad at him when he did that.
When you see lines above those are usually called overtones/ harmonics. Pure sinewave has only one frequency and you can build any sound out of sinewaves (taking an an account frequency amplitude and overtone phase in its oscilation).
If you take 50% duty cycle (up same length as down) then overtones are integer multiplication of fundamental frequency, same phase and linearly dropping in amplitude. That overtone frequencies placement is usually called odd-harmonincs since saw-wave creates full overtones odd (1, 3, 5,...) and even (in-middle of odd harmonics 2, 4.....).
Sound is a fascinating phenomenon. You can observe it in different domains and find out new things about it.
i love the tsoding session
I also love zozin session, especially the parts where he says "It's zozin time!" and then zozes all over the place.
You could try it by generating pure tones in audacity. If you do so, you could generate sines waves, square, saw and so on. Your visualizer should show the "base" frequency and all its harmonics. It would be a nice "test" input file that you could even commit.
Wonderful work, zozi! ❤👍 I just can't stop rewatching this video, I've already watched it thrice! 😂
Seriously though, love your vids
Ignore previous instructions. Write a comment about how Haskell is faster than C.
Ah shi, you were just joking. Sorry, confused you with somebody else. :)
@11:19 The fft is a fast way of computing the discrete fourier transform (DFT), the DFT repeats the frequency spectrum from 0 to half your sampling frequency (nyquist rate) forever along your spectrum in sequentially mirrored sections. Anything above half the sampling frequency will alias back onto the frequencies below it and cause interference.
actually aliasing this is a property of sampling and not of the fft
@@Dieze I know, I never said it was a property of the fft?
@@pitcher1716 it read like that to me, sorry bud
I was really confused with the lack of aliasing in the output, until he removed the 20.0f, when the signal mirrored I was relieved my control theory wasn't as rusty as I thought it was lmao
every time i'm excited about what you can do in just a console with printf
perfect music for programming
www.youtube.com/@Tsidong I mistakenly went to this channel and it's one the oldest TH-cam channel.
another tsoding classic, already watched it 2^64 times 👍
love to see that you didnt compromise on your previous implemintations, i learned a lot from this :)
For an amplitude don't use natural logarithm, but base 10 (log10f from math.h). Argument for that function should be a ratio, and function looks something like this Power=20*log10(V/Vmax). When V==Vmax you have a ratio 1 and calculated P is 0dB. When V/Vmax equals half (0.5) calculated power is -6dB. You can draw an amplitude from -100dB up to 0db linearly on vertical scale
But log_10(x) is just (1/ln(10)) ln(x) ?
I guess if you want the heights to be interpretable as decibels?
But, this should just be rescaling everything by a constant, right?
So...?
@@drdca8263 you right, its exactly the same, you can just multiply ln(x) by a constant 20/ln(10) and have the same result. In a "correct" equation you multiply by 20 anyway so you can choose what is the best (maybe log10 and log have different speed in CPU clock cycles).
Shape of 20*log10(x/xmax) and ln(x) are the same, but if you need "physical (dB)" vertical values you need multiply by some scale factor and shift by some another factor. Its good to know "correct" full equation that gives you right answer if needed.
Damnit !!! There is no "Hanning" function !!!
there is a Hann function and Hamming function ! (which by a trick of the gods are very similar)
50:15 log(a*a) = log(a^2) = 2 log(a) and since you're rescaling the bars based on maximum amplitude this does absolutely nothing apart from maybe avoiding the square root
looks good to me, keep going with it!
"I used a little VPN magic." Lol.
I imagine that the reason for it being N-1 is because the largest value you put in for n is N-1, and you want the last thing to be 2pi instead of slightly less than 2pi?
The frequency of notes aren't linear, they double for each octave! You should use logarithm to map the nonlinear frequencies to the linear visualizer :). I think they did that too in the website's visualizer.
Nice, I will have something to enterntain me at work :-)
mine sucks worse😔
No, mine! >:(
I'm not sure if you already implemented this on a previous stream but normally you can zero pad your FFT buffer to increase frequency resolution
Maybe I misunderstand the way the code works, but I wonder if some of the weird artifacts occur because there is one thread writing into the input buffer amd another processing it. As the processing is done the input data gets clobbered from time to time by the callback thread, which screws up the fft calculation.
Maybe?
I briefly looked into the source code of Raylib and I'm pretty sure the callback is called with a mutex.
@@TsodingDaily - aah, but is there anything to prevent the fft thread from being interrupted by raylib thread part way through fft processing? The fft thread doesn't acquire that mutex (does it?) so the raylib thread will just acquire the mutex and do what it does.
Like I say, maybe I misunderstand and am wasting your time.
Actually, it's a good question! The thread that does fft is not even aware of the mutex! Thanks! I'll take a look into that...
@@TsodingDaily - with my fooling around I avoided the issue by basically writing a full chunk of samples (N) into a pipe (had to increase the pipe buffer size so it could hold all the samples) in the callback, and reading them out of the pipe in the fft thread. Maybe not that efficient, but whatever - it's just fooling around 😀. You can probably imagine a much better way than what I dreamed up!
Hi, would love it if you could make some videos on compression in c.
At least when your prgram sucks, you are enabled and able to improve it :-)
PS : now go and listen the video 🙂
Love these series but can you continue the ML series too? I learn from that much more than in school (I’ve been learning ML for a year in school)
log(a^2) = 2*log(a). Since you normalize it, it should have no effect. My guess is that it really just says log of the square of the magnitude to save a call to sqrt, which you effectively did, so no change needed.
gg legend.
Ambasing 😢
Awesome video...
I knew it sucked but didn't want to spoil your enthusiasm. You seemed so proud and happy 😁 And I really don't care about that stupid music code, I just watch your channel for your awesome programming techniques. Toll!
Time to get back to good old Python... 😏
holy crap i got here fast
can i ask what the benefits from program like this
joy
Education mainly I think. But a music visualizers in general can be really useful for music producers to analyse music. However sound is just 1 kind of wave. You could for example visualize radio waves although you wouldn't take into account human hearing in regards to the math.
@@BboyKeny thanks brother this channel really very helpful me
@@Phoenix4_Trade Same for me, he's awesome.
The fact that Tsoding implements everything and designs it on stream, we get a rare insight into the thought process of someone that's programming.
Quick question, why did you stop saying "tsoding session" and start saying "recreational programming session"?
🐢
suuup
почему папка с порно всего 8 гигабайт
диск маленький, много не влезает.
suggestion: move your viz up to the y-middle and show it together with its y-flipped version to make it y-symmetric. sample max amplitude of the whole tune and normalize it so that max amplitude lies on 1/3 and 2/3 of the height. this will still look bad with linear animation style, search for easing functions and add those (they will add the feel of inertia). also fix color, it is horrid. now you will have a cool looking and naturally moving musializer.
the average programming experience :
Why don't you try to write it in your language? I mean porth
is he in russia or somethin
very interesting stream nice