Deep Learning Techniques for Music Generation – A Survey | Hacker News Reader

Deep Learning Techniques for Music Generation – A Survey

arxiv.org

96 pointstim_sw26 comments

There is even now an international research workshop on deep learning for music that was created in 2017: http://dorienherremans.com/dlm2017/

Also a good overview (albeit, less detailed): https://medium.com/artists-and-machine-intelligence/neural-n...

Would a computer endlessly churning out music on par with Beethoven, Chopin, and the Beatles make us happier as humans?

Some of us, yes, some of us, no?

Shall I refine the question then?

Would Beethoven, Chopin, or the Beatles be happy if a computer churned out equivalent masterpieces?

Define equivalent? I'm sure they had competition in their day.

The best answer I've heard to this and your previous question was at a QandA with Dr. Daniel J. Levitin. His answer was to a question "Is there any music that can make everyone upset or unhappy?" His answer was that in all his research the only music that can make someone upset or unhappy, was music the listener didn't like. It's 100 percent listener preference. The same was true for music a person did like. Unfortunately, the only way we could find out if Beethoven, Chopin, or the Beatles like the music is to ask them. I'm sure we'll find both support and disdain from the music community on AI generated music. Time will tell if there is a major preference on either side.

Dr. Levitin has some really interesting research on music, but also has some great easier to read books for those of us without a neuroscience degree. My favorite is called this is your brain on music. The world in 6 songs was also a great read.

http://daniellevitin.com/publicpage/

I wonder how relevant this perspective is to the discussion going on at [1]. Particularly, I am curious how people interpret the notion that a deep learning system could predict a person's musical preferences by their facial features.

[1] https://news.ycombinator.com/item?id=15219906

Also, there are patents available for deterministic models based off of neurophysiological response data obtained from various sensors (primarily heart/pulse, galvanic capacitance of the skin, body temperature).

One model mentioned in the patent uses Nigel Osborne's INRM (Innate Neuro-physiological Response to Music) paradigm to make predictions.

"Also?" I was only bringing the subject up for discussion - I think it would be ridiculous to suggest that facial features could actually predict a person's musical tastes. These tastes change in some people from day to day or even minute to minute. People also subjectively evaluate music very differently, even when it's music they enjoy listening to.

As for those patents, I would caution against interpreting them to mean a person's musical preferences can be inferred or exploited in this way. Merely evoking a physical response is not the same as satisfying a person's preference or desire. Particularly, a physical response can happen in the absence of a person's consent to receive the stimulus, and in rape trials this is legally recognized to exclude evidence of such a physical response from being used by the defense.

That's not refining the question, that's completely redefining it.

We don't need to churn out more of it, we can't raise to the level of the current supply of music. What we need is more power to appreciate genres of music, not more music.

Have you tried researching classical music on YT? I have done that intensely, for about 7 years, and I can tell you that I didn't get to listen to more than 1%. There are thousands of composers with hundreds of thousands of pieces and millions of interpretations.

If we create computer music it will be done to expand the space of creativity, as usual. Classical music is particularly daring in breaking its molds and expanding to new forms.

My favorite imagination game is to pretend I am listening a contemporary piece together with Mozart, Bach, or Beethoven and watching their reactions. So I am doing in my head a simulation of the old masters, like AI is doing, except that I am trying to listen from their perspective instead of trying to create more music in their style.

i don't get this at all. we need music that is about who we are now, and what new sounds we can make and imply and expect and imagine. technology has changed that just like the modern violin bow did for romantic era violinists, computers did for Varese, Antheil, and Percy Grainger, autotune did for Cher, hammond b3 did for jazz, trains did for Honnegger, race riots and loopers did for Strve Reich and on and on... we are at a different launch pad with ML, one that might allow us to shift the focus of our musical labor and practice, one that might allow more diversity in just who gets to make music (hopefully more people who would not otherwise afford to, for one)and more openness in what is called music, in what we consider to be legit instruments. why can't we use all sounds more accessibly like instruments? Why can't we hear how children might hear (usually because they have not yet learned technique) -perhaps our whole concept of what sound and music actually means could change... but first we have to wear ourselves out with all the Bach/Taylor Swift/Ligeti hybrids until we see how shallow that feels to be trapped in that space. Then, a new concept will emerge. I think it's machine learning that might actually get us there- to the end of justified by labor where the real art is. I think we should face the fear we seem to have that we might not actually have ideas--- that's where the art is. Not in a Miles Davis/Mozart hybrid.

Your message was interesting and I agree with it up to the point where you say we could be trapped by Bach. There is enough depth in Bach to still fascinate us 300 years later, if it doesn't it is a lack on our perception ability.

On the other hand, I wholly agree with innovation in composition - Bach, Mozart and Beethoven (+ many others) excelled in innovation and were sometimes considered downright weird for their time. Even Chopin was considered weird and his music is the easiest to listen to.

As composers, is is/was their duty to be innovative, but us, as listeners, should try to assimilate all periods and styles, and the old ones have a depth of feeling that is rare today.

perhaps we should be less passive listeners. Bach was just a dude who was great at counterpoint and his job. we aren't trapped by Bach- we are not thinking creatively if we think stylistic hybrids is artistic thinking. Shoot -- Bach was doing that his dang self. The point is-- art isn't something that should be justified by labor. It sure is a lot of labor to write a fuge in counterpoint like Bach did. But only if you are doing the math by hand. Time to get to the end of labor so we can see through it to an actual artistic idea about ourselves. The passive listener might be a great place to start.Bach was not a storyteller. He was into absolute music-- mathematical purity-- a marriage of form and function under a set of rules called counterpoint both laterally and vertically-- the ultimate musical 3d Sudoku meets Rubik's cube meets Go meets chess-- in sonic form. What you hear in it as emotion is what you see in the surface of rippling water, and that is kinda the point of its technical and formal strictness and purity. It is musical math. Math has no emotion-- you do. Bach aimed for this kind of pure dedication to musical math. But it isn't idea. It really isn't. But there is something compelling about that which is concrete giving rise to our own ideas. I guess the poet William Carlos Williams was into the same thing-- bringing out the concrete so something personal can float up from that image- so it can be somehow universal. I guess that is cool. But as a lifelong performer of Bach, This Bach worship is its own religion. They didn't have cameras or computers or safe surgery then. (bach dying of an infection from cateract surgery -- knew nothing of blood type also)But we DO! And I think it is time to move on from the whole "wow, check out that perfect form and counterpoint he did by hand" kind of thing. It is a great player who can give real life to pure musical math, and that is what Bach did for us-- he gave us a nice, sturdy, non-flaky canvass, and we all appreciate it. But time to move beyond the hand labor being recognized as the art in music composition and music performance too (wow, look how in tune! how fast! how well structured in phrase! how expertly varied in timbre and sonic texture! how emotional!) It really isn't a marvel of anything but hours of time spent doing counterpoint or scales or technique. None of us expect to be lauded for our facility at speaking our native language-- or even a foreign language. It is the same thing, I promise. And hopefully, machine learning will show us this sooner rather than later so we can get into some new artistic idea. And yeah, loads of performing musicians disagree with me on this. But hey- loads of peoole think they were abducted by aliens too.

It depends on what you mean by "on par".

If you mean "endlessly churning out songs that sound pretty much like Beethoven, Chopin, or the Beatles", then probably not. Humans do that too.

If you mean "endlessly creating new, innovative, and enjoyable material, like Beethoven, Chopin, and the Beatles did", then I think it's a very interesting proposition and lots of people would be happy to have that. I am skeptical it can be done with ML though.

I think it would be a huge step forward for music or any artform, just like photography was for visual art. after the camera, we could relax on the technique of copying things as we see them and move on to ideas. perhaps ML could help human music finally get real about ideas. It could be wonderful to be free of the labor so we could really focus on the other aspects of music this kind of development could illuminate by making the formulaic nature understandable by everyone, driving a new interest in a frontier in music we were not free to explore while chained to techniques that are just techniques and not actual mysical ideas. I think music might more emphasize connection between people and a given moment between those very specific people. It might bring the former audience on par with the composer and as a composer, i'd like to see that in my lifetime!

Not everything needs to be created for happiness maximization. It'd just be really really cool.

Maybe its not going to be about completely copying artists. Just attributes from them. What if training the models becomes the art form itself and datasets are the building blocks you use to compose with. Imagine 'composing' models by selecting the training sets. Example: train with the rhythm style of James Brown & the tonalharmony of 50% Bach & 50% Miles Davis.

Focusing on happiness, what if the computer were just endlessly working on making humans happy? Would that be any better?

You might end up with this: https://www.youtube.com/watch?v=Isrd7E5nzIQ

Or perhaps something much, much worse like electronic crack.

You could argue that we've already achieved that: http://nypost.com/2016/08/27/its-digital-heroin-how-screens-...

Humans are flexible and adaptable. We can take any good fortune and make it ours while feeling entitled about it. In one generation or two, the new normal will be old. I think my parent's generation has seen a more dramatic raise in their conditions of life than I ever will (going from countryside to the city, in the period around 1970's). My grandma's never had a water closet in their life and raised water by the bucket from the well, and yet lived well above 80 years old, yet I now use a Japanese washlet - spoiled rotten, that's what I am by her standards. I don't even have to wipe my own damn ass.

I'm a simple man. Four chords are enough to make me happy.

https://www.youtube.com/watch?v=5pidokakU4I

You can experiment with different progressions at this CodePen:

https://codepen.io/jakealbaugh/full/qNrZyw

Maybe it's not Beethoven or Chopin, but if it's good enough to compare with modern pop songs, then there should be a market for it.

Depends on the level of creativity and if it can break out of the learned styles and create something different.

Different is easy. Different and generally pleasing (making sense) is a much harder proposition.

ML models have problems handling even medium level structure much less something long term.

it has existed for many year of generate music with ai.

Yes, that's the only way it could be a survey.

Reply on news.ycombinator.com