Hacker News Reader

Comment by 323 | Hacker News Reader

In Go, and Chess, there are a number of "rules": you should never do this (move the same chess piece twice in the opening), you should do that, ... And then AlphaGo appeared and dismissed all this and did just the right thing for the particular game being played. Know the rules, but if you are an expert you can break the rules.

I wonder if AI will do the same thing in music, it will use the "perfect" tuning suitable for a particular piece of music and dismiss this idea of a universal tuning scale.

View on HN

You can't realistically have a different guitar or a different saxophone for each and every piece you want to play, and those frets and holes can't be freely moved around. It gets that much worse when you consider "installation" instruments like carillons or pipe organs.

AI just literally, fundamentally can't "dismiss the idea of a universal tuning scale", because whatever per-piece optimisations it can come up with still need to be realised by physical instruments at some point. The idea of a good-enough compromise solution that allows you to play a wide variety of pieces on a single instrument is just too damn important.

There is more to music than just physical instruments.

In popular commercial music you do literally have a different instrument (synth setup) for each song.

But even if we talk about guitars and saxophones, I was speaking about AIs which directly output a sound file, not a music sheet. So they can synthesize a fake saxophone which is tuned in a weird non-physically possible way, as if each note was played by a different physical saxophone that the musician switches to.

You specifically brought up Alpha Go dismissing the conventional wisdom on how Go should be played. Many of the things we thought we knew about the game turned out to be wrong and the game as a whole was turned on its head.

None of that applies to music. Nobody who studies this stuff seriously is under any sort of illusion that 12-TET is the "right" way to play music. I know a fair few professional musicians, and I've "talked shop" with as many of them as I could, and the deficiencies of 12-TET recurringly come up. There is nothing here to "dismiss".

Don't get me wrong: The idea of computationally-optimised tuning sounds really interesting, and the discussion of what we should be optimising for would itself be fascinating to follow. It's just that people are already doing that sort of thing manually today, so there's no big "oh no we're doing it wrong" dismissal of the status quo waiting at the end.

> None of that applies to music.

But how would we know that? People thought music was figured out and then atonal music was invented/discovered/re-discovered (whatever you prefer).

We are somewhat speaking about different things. You talk about people playing instruments, and you are sort of right, all possibilities were explored.

I'm talking about audio files with songs, many of which are currently being produced with software using a specific tuning (typically 12-TET). But in this world the tuning is just an artifact of the production process, it's not fundamental like in your world.

The current picture producing AIs don't start with a blank digital canvas and drag digital brushes over it, they synthesize the image in a holistic way and in this world the "brush" can be unique at each position.

More precisely, I'm thinking that music producing AIs could make music where the first 5 seconds of the lead instrument uses 12-TET and then switches to another, the backing bass track uses a different tuning, the vocal sings to yet another one yet it all comes together beautifully. And the tunings used could morph during the song duration. In a way this means that there is no tuning at all.

I think the key difference is that playing go is about winning (at least, presumably that's what the AI is optimized for). Music is not.

(I also agree with others in this thread that the popular commitment to equal temperament is exaggerated -- it's not all that uncommon to hear good musicians of various styles playing/singing/synthesizing "out of tune" music for various effects).

Again, the point isn't that there's nothing left to learn. There's plenty to learn, and plenty to explore, and the whole field of applying computational methods of all sorts to music is a treasure trove waiting to be explored.

What I'm saying is that the situation with Go was completely different. The Go community was utterly convinced that the state of the art was within a couple stones of the hand of god, and Alpha Go thoroughly disabused them of that notion. The status quo was completely shattered, and the community's understanding of the game as a whole was completely upended. It's entirely fair to describe the situation as "and then AlphaGo appeared and dismissed all this"

The situation in music is very different. Ethnomusicology has been a thing since the mid-20th century, and musicology in general has swung away from prescriptivism and more towards descriptivism. There can be no earth-shattering revelations here, not because our current understanding of music is unassailable, but simply because there is no earth to shatter to begin with. AI-drive computational music might produce some innovative work around how we understand pitch and tunings, but that work won't dismiss our current understanding of those things, it'll sit alongside it.

Ok, this is fairly long winded, but the point is that I take issue with the "dismiss" part of it all, I guess.

> You can't realistically have a different guitar or a different saxophone for each and every piece you want to play

Looks quizzically at 44.1kHz-u16 audio sink.

Pretty sure I can, actually; my computer's speakers certainly do, barring a rare handful of groups of songs that were recorded at the same time and place.

People broke conventional rules with success in both Go and chess before AlphaGo and AlphaZero.

In a similar way, people have been using particular tunings for their songs for a long long long time. The idea of a universal tuning scale is relatively new. No need for AI to point us away from it, we already did that ourselves.

Yes and it was common for the player to tune (and presumably retune) fixed-intonation instruments like the clavichord... Bach famously had his own tunings that "sounded good". You still see this today with e.g. hurdy gurdy players who are constantly tweaking their tangents by ear.

Your post reminded me of a video[1] Adam Neely made where he explored some (to me) weird tunings, starting with one where A = 432Hz.

As someone who hasn't taken any musical theory or similar, it was quite interesting to hear.

[1]: https://www.youtube.com/watch?v=ghUs-84NAAU

>where he explored some (to me) weird tunings, starting with one where A = 432Hz

That's just changing a convention, not a tuning in the sense talked elsewhere in this thread (how we divide notes), but "what our starting frequency is".

A=432 and A=440 is just as arbitrary. They just had to pick something so they would all match.

The main difference is that 432 is associated with a set of new age, healing, "universe", etc. BS claims in certain "spiritual" circles...

The one special thing about A=440 is that it is international law, as defined by the treaty of Versailles (yes, the one that ended World War I)

What's the penalty if you make an instrument tuned to A=442? Do you get dragged to the Hague International Criminal Court?

Several orchestras use different A's around 440, and nobody is getting prosecuted. 441 and 442 are popular right now, although some go as low as 438.

In baroque music, ~430 and 415 are also very common since they are thought to be the historical pitches of "A"

Friends recorded this album --https://alisonperkinsandnicolasbrown.bandcamp.com/album/all-... -- with A somewhere in the neighborhood of 360Hz.

360Hz?

That's so low, it's more like playing the piece a three semitones lower than an alternate choice for A4.

If the piece was in A, it would be more like playing it in F# (while still using A=440).

(Of course if you did that, the "sweet-spots of 12TET and its off-notes would be different, than if you played with A=360)

Some late baroque-period harpsichords had a selectable A: you could chose ~430, ~410, or ~390. The adjustment came from sliding the keyboards to the left or right based on which A you wanted. Supposedly A = 390 or even lower was used by the French in the renaissance, so you wanted your harpsichord to be able to accurately play historical music.

None that I'm aware of.

Tuning your orchestra high was sort of the 19th century equivalent of the modern loudness war. The problem is that orchestras tuning to ever higher pitches meant that singers had to sing higher to match, and it was putting serious strain on their voices, which can easily lead to injuries.

Having some sort of agreement setting a standard was just something of an "enough is enough" sort of moment. It just amuses me to no end that this was achieved by writing it into the Treaty of Versailles, of all things. We're settling a freaking world war, so let's make sure we settle the issue of orchestra pitch as part of the treaty.

Though it is true that a lot of older string instruments weren't designed to take the tension of modern strings at modern pitch, and some of them really open up at a slightly lower pitch. I'm building lyres, and many lyre people are from that "A432 resonates with the universe" crowd, so I've been using it — and I can't deny that there seems to be a sweet spot in sound for a lot of instruments at that pitch. I honestly have wondered if there's some physiological reason so many people prefer it.

Aside from the universe, there's a very practical point related to this, which is that the instrument has other resonances besides those of the strings. E.g. I have read that the frames of harpsichords are tuned to particular resonances, which is part of what gives different keys different qualities.

Well he does go into Pythagorean tuning later in the video[1], both a proper one and one which was made to "look nice", so bit more to it than that from what I understood.

Or I might be wrong, I know nothing about music[2].

[1]: https://www.youtube.com/watch?v=ghUs-84NAAU&t=517s

[2]: https://www.youtube.com/watch?v=s6EaoPMANQM

What would the AI's reward function be?

Song gets into Billboard Top 100? Song view count on YouTube?

And for earlier stages you can have human raters or similarity with past successful songs.

Feedback from a neural link which indicates satisfaction in the listener?

It won’t get terminated on AWS.

Reply on news.ycombinator.com