Lantertronics - Aaron Lanterman
Lantertronics - Aaron Lanterman
Hi there! My name is Aaron Lanterman. I am a Professor of Electrical and Computer Engineering at Georgia Tech. When Tech switched to "distance learning" in March 2020, I decided to put the lecture videos I was creating for my classes here in the hopes that some folks outside of GT might find them useful. We'll see how things go from there.
The opinions expressed here are my own, and do not in any way represent official positions of the Georgia Institute of Technology or any group within it.
Пікірлер
The modified FM Timbre is really much smoother to control, Thank you for introducing us! But it is also kind of lacking in the harmonics - from a musical point of view at least - is there a way to remove the "bessel-swirl" that does not remove as much spectral content?
it's not Bjork it's Fever Ray
I think in all these graphs I'm hearing changes above 3000 that aren't visible on the graph. The spectrogram suggests the sound mostly isn't changing, but ears say otherwise. At my age I trust my eyes more than my ears, usually, though...
Another fascinating video. I added "spectrogram" to my TODO list.
Thank you, Prof. Lanterman. This series on FM variants is amazing!
I was thinking, a series on the Electro-Smith Daisy Seed would be phenomenal. It's a great tool for DSP.
Euler lives! e^iθ = cos(θ) + i sin(θ). The imaginary sines and cosines are just the hyperbolic functions cosh and sinh. I love Bessel functions.
Regarding AI music generally, it makes me think about a few pubs and bars here (and it won't be limited to them) that decided to ditch live bands and replace them with DJ's to save money. It was a disaster for them because the clientele changed from people who would buy drinks from the bar to pill heads who barely drank water. Now they're trying to claw back the live bands. I feel the same thing will happen in the music industry with AI. The faceless people in charge will want to remove the expense of living talent and replace it with something cheaper they can also control. At the same time, no one will pay money for the soulless music produced. Of course, the industry will have to try it first before they realise it's a huge mistake.
So much AI music reminds me of the filter tracks on late 90s CDs, where you'd hear a couple of good songs on the radio, and buy the album, and it turned out those were the only good songs on the album, but that would be the only way to buy the songs. Like, it's passable background music but doesn't hold up to any scrutiny. And I think there's a limit to how much it will get better -- the algorithms may get more clever, but ultimately yo need exponential increases in the size of the training data to get linear improvements (it's sort of running on the opposite of Moore's law, even if the underlying computers are subject to Moore's law).
the tiptop buchla dont eally sound like a vintage buchla so neither will the behringer. but it will be option for people who wanna mess withat westcoast design. all those behringer clones are there for people who are not purist or oldschool high end collectors. you can buy a good sounding analoge synth for 90s backyard market prices. its a good thing with all that digital generic music to inspire people to make something original. a lot of the sonic quality is processing so i can make a behringer sound like whatever oldschool synth soundif i process it right. a 909 is not that special unless you add those pedals little tubes or whatever to get that sound. it was never a plain synth output in most cases of those iconic sounds.
Why do you think TipTop Buchla sounds different from vintage Buchla? Do you think TopTop Buchla also sounds different from modern 4U Buchla?
The first of the “Nick Cave” ones very much reminded me of Pete Murphy. I wasn’t surprised to see that someone beat me to it in the comments section!
6:07 I burst out laughing
thank you for sharing this
i listen to most of these artists, björk and david sylvain in particular, and can't say THAT'S IT!!! about any of them except for the most egregious johnny cash example. there's still very serious questions about its training data, but every time i start reaching to hear the real people in these examples, it sets off a lot of cognitive bias alarms in my head. as humans we have a tendency to vastly overrate our own pattern recognition systems, and to try to sort the unfamiliar as adjacent-to the familiar until it's a blatant splash of cold water to the face.
Another couple of years and synths will be at least virtually "manufactured" by a second grade chat bot.
The Leonard Cohen track (LC1) sounds a bit like Bill Callahan to me
Aurora's Veil sounds a lot like Emiliana Torrini, also Icelandic.
As does Twilight Echos
I bet it would do Don McLean
Rename it “The Plagiarizer”.
the instrumentals at 2:20 reminded me of one Bjork song, army of me. It seems like a faster version of its intro. Anyway, great video as usual Aaron! very informative, ty for making it.
DS2 sounded like Bowie !
David Bowie and Nick Cave is what i hear.
The “Nick Cave” songs sounded more like Peter Murphy (Bauhaus) to me. The “Johnny Cash” tracks were downright creepy in how close they were. The “David Sylvian” tracks had some similar elements, but I not enough to my ear to say they were him.
a lot of those voices are very similar to the artists you were going for. but maybe we are just learning that the space of all interesting voices isn't that big. like, bjork, the identifying part of her voice being her voice in that snippet is the icelandic accent, but there are a lot of girls who have an icelandic accent. it's similar to how you can only have so many shapes of fish and so many forms of sports cars before things start looking like each other. so what are we talking about here, really?
In Bjork's case there's a distinctive timbre and phrasing that goes beyond just the Icelandic accent; I was able to pronounce also sorts of female Icelandic vocals that weren't distinctively Bjork. In fact I suspect that Udio changed their algorithm since the lawsuit hit, since now using the same prompt I used a couple weeks ago tends to produce less Bjorky samples. But your final question is a good one, and the answer, for the moment, is that I'm not entirely sure. ;) You hit on something deep with the question of the "space of all interesting voices." I'd like to explore the underlying latent variable space that something like Udio winds up working with. Just what is the dimensionality of the human voice?
@@Lantertronics the study of novelty, in general, is very difficult and fraught with self-deception. i think the only thing that i saw make actual progress in that realm is assembly theory, a recent paper by Sharma e.a.
@@Lantertronics on top of what i replied earlier, i think to fully understand how much novelty there is to be had, you should talk to psychologists and sociologists about how our concept of novelty is formed and how it is used. How many voices do we need to be able to tell apart and why? How is that used in social situations, how was it used historically when our brains were forming? That could lead to some good answers. Also another thing. I think popular singers often occupy a niche that they did not create, but one which they represent very well. For example, johnny cash is a poster child for "generic old man voice". If you imagine what an old guy sounds like, that's probably going to be close to his voice. Britney Spears had a generic teenager voice. Eminem has a generic misspent youth voice. I think *that* is why they become so iconic: the icon existed before they did, and they just fell into it. Lucky more than anything. This whole observation touches on my earlier statement that you can only have so many interesting voices.
David Bowie
Weirdly, every time I tried to *intentionally* get David Bowie, I wound up with something that sounded like the Beatles! (In the case of these examples I was going for Tom Waits).
I just tried the exact same prompts you used for your Bjork result and got nothing remotely close
Me too. I’m wondering if the lawsuit prompted Udio to change their algorithms to try to be less obvious. I think in the public facing interface they have a post processing step that checks to see if it resembles copyrighted content and rejects it if so; maybe they changed the threshold for that?
The one thing that really struck me about all of those snippets is how they all failed to move me.
AI output has the feel of filler tracks from late 90s CDs when there would be two good songs on the radio and they’d be the only good songs on the CD, but you had to buy the entire CD to get those two songs.
TBF, 99% of music made by humans has the same (non) effect on me.
6:54 was 100% Waylon
I wasn't familiar with Waylon Jennings, but I just went and listened to some, and, uhm -- wow, you are right!!!!
What if AI revenue had 100% tax?
I think you could make an argument for an X-percent tax, where you figure out X by determining what portion of the value is from the underlying algorithm -- the particular clever tech that Udio or Suno or whoever created -- and what portion of the value is from the training data.
They should stick to ripping off the Black Eyed Peas and Three Doors Down, then nobody would find any of it distinctive enough to show their hands.
The first part of the Leonard Cohen sounds more like Eddie Vedder to me.
Waylon!!!
Wasn't really familiar with Waylon Jennings, so I just listened to some after reading your comment, and yeah, you are 100% on target on my 2nd "Johnny Cash" example.
That is David Bowie, however, I heard some Mark Knopfler and Tom Waits in the first one you played. All the Icelandic sounds like a mix of Björk and the Faroese singer Eivör Pálsdóttir - Björk has a stronger, more commanding tone. The Icelandic pronunciation is excellent except in the first one where it sings "Falla" (e. falling) - this is pronounced something like "Faddla" normally; but Björk would spice it up with extra icelandic aspirated 'h' : "Fahhddla"
For a fact, I never thought of gain as anything. It was like a black box. I started playing about 25 years ago. A masters in music is one of my three advanced degrees. I had a daw early on so maybe I’m not a “guitar player”. I knew that clipping was bad and good. I knew the gain knob made it louder and more distorted. A gain boost boosts gain, “boost pedal”. Other boost pedals include OD, distortion, and an EQ pedal (see Kerry King). An EQ can be a boost, but a boost generally is not an EQ, unlike an OD or distortion which both either necessarily or optionally offer some tone sculpting.
Wow.
Thanks! :)
I was so excited with this that I made an approximation of Config #02 on the Gotharman's Urano modular-in-a-box with either wavefolding at the end, or having oscs 1 and 3 be wavetable oscs with VCAs 3 and 6 modulating the tables. I gets quite weird (and fun). There is a massive amount of sounds to be found in just this one configuration/algorithm. Need to try out some of the other Configs.
I hadn't heard of the Urano, it looks really cool!
@@Lantertronics It is discontinued, but his Zaturn continues the tradition.
M4 sounded like adjusting a slider between Cohen and Waits. M1 was more Bowie doing a Waits impression. M2/3 sound a bit of David Sylvain with an American accent
A bit Jonny Cash there, but not quite like him, could be Tom Waits, but he's it neither. Could be somebody I don't know.
David Sylvian, the resemblance in the third one is uncanny
Guessed Tom Waits immediately.
So, who is it supposed to be? You never said.
I need to make a follow up video on this. I used tags from rateyourmusic to try to get Tom Waits, although I don't think this is all Tom Waits, I think Udio is mixing in other vocalists here.
This wasn't as clear cut as my Bjork examples.
Tom Waits
Quite obviously Tom Waits.
Sounds like Tom Waits
1 chris rea
Billy Gibbons
It's all Frank Farian in disguise
Hah! "There was a man from Russia long ago..."
My bet is Tom Waits.
Tom Waits as fuck hahaha wow, its Tom Waits if he got hit in the head with a sack full of old brass doorknobs before writing the lyrics.
All artists have "trained themselves" on music that came before them. There is nothing wrong with AI doing the same.
This is common misunderstanding about how these algorithms actually work -- people hear "neural network," and say computational neural networks were inspired by our neural networks in our brains, and start assuming equivalences than aren't actually there. The way these algorithms are trained and represent knowledge and create their output is fundamentally different form how humans do it. Ask yourself why AI image generation algorithms will reproduce the Getty watermark when no human would do that, and you'll start down the path to seeing what's going on. One can be inspired by Bjork's vocals, but no amount of inspiration and training will actually give you Bjork's timbre. You'd need to transplant her vocal cords for that. This gets incredibly complicated, and I wouldn't except you to agree with me with just a few paragraphs. I plan on addressing these issues in detail in a future video.
@@Lantertronics That wasn't what I meant, but I get where you're coming from. That is the technical aspect of it. My angle is basically that I don't accept the right of the recording industry to make money off AI models using the available music catalog. I want AI models to be able to freely train themselves on all available music.
@@Junyo I could sign on to something like that if the resulting AI model itself is also free -- although I don't see Udio signing onto something like that. ;)
mix of tom waits, bob dylan, and david sylvian
M4: mark lanegan M1: Nick Cave/Bowie M2: Nick Cave M3: The lyrics and tune sounds like a Charles Manson folk song