Comment by exDM69 | Hacker News Reader

Comment by exDM69 | Hacker News Reader

I predict that streaming, listener analytics and metrics will have a much bigger effect on music production by removing the human producers for mass consumption pop music in the next two decades. The writing is on the wall.

We're already at a point where the producers (who compose the songs and usually even create the audio using computer software) spend a lot of time analyzing the metrics they get from Spotify et al. They look for stats on how many seconds someone listens to a track before skipping to the next one, do they turn the volume up or down, which songs are paired to the same playlists, etc etc.

When reading interviews of musicians, they frequently bring this up and usually not in a positive light. You can also hear the results if you're exposed to modern pop music - when there's a hit single, there will be quite a few copycats that will shamelessly borrow the elements of the hit. This is not a new thing but the degree to which it's occuring is.

Now what's good at looking at metrics, doing tiny adjustments and gradient descending towards some local optimum and cheaper than hiring a musician? A machine learning algorithm of course.

While it's not common yet, there have been some fairly impressive demonstrations I've heard. Computer-generated music notation is easy to generate (read: number of bits in the output is rather low) and has been experimented with for a long time.

But this can go even a step further, rather than producing sheet music, an algorithm can generate the audio directly. As in the PCM audio waveforms (or maybe spectrograms, frequency domain might be easier for algorithms). I recall hearing AI-generated classical music audio (found through HN) which sounded fairly convincing to me (albeit with a white-ish noise on the background).

If you're a professional pop music producer, make the most out of your career while you can. You're closer to being replaced by robots than you think.

I'm not suggesting that human musicians will go away any time soon, I'm sure people will enjoy playing instruments with their own hands in the future. But the kind of producers that create music with a computer mimicking popular songs and tracking analytics from listeners will be, if not replaced, at least heavily assisted by algorithms more and more in the near future.

The tools evolve and that's a great thing.

Music, as painting, drawing, sculpting, writing, stays a mean for people to express themselves, to share their feelings & thoughts.

But it's also a mean to grab attention and manipulate it, too.

> I'm not suggesting that human musicians will go away any time soon

It already happened. Since records exist. See the quantity of shows where there's only a voice performer alone on the stage or on the street, with pre-recorded music.

Not that it is a bad thing - it is a _different one_ that does not remove anything, only adds. Where the purpose is not to perform, but to grab attention with less logistics.

Live musicians are still a (great) thing, but it's definitely not the same thing.

I agree with your main points, even though I'm unsure of the time scale.

However:

> But this can go even a step further, rather than producing sheet music, an algorithm can generate the audio directly.

Even if this is a cool idea, I don't think it ever will be deployed large scale. Good quality beats/samples and increasingly even software instruments are becoming cheap. Working with raw audio seem like the wrong abstraction layer for music, almost like working with pixels instead of characters to synthesise text. While limiting - working with raw audio can possibly create completely new sounds - it will probably always be cheaper and yield better results to work with some representation of the music instead.

> Working with raw audio seem like the wrong abstraction layer for music

I do kinda agree with you, but I wish I could find the classical music audio clip I mentioned. It was generating a fairly convincing clip of audio complete with an orchestral backing and a piano solo track. Maybe that would change your mind a little.

Unlike text which is discrete characters, music is a continuum of pitch, tone and other characteristics.

Maybe it won't be generating the music, but adjusting the tone instead. There's quite a lot of effort in mixing and mastering music, which sounds like good territory for ML.

To return to your text analogy - it's more like generating convincing images of hand-written text. You might generate the discrete characters one by one, and then proceed to turn that into pixels with another algorithm.

The big advantage of using raw audio is that there's an abundance of data (all the music that's ever been released pretty easily scraped from the internet), whereas "semantic" data is rare, expensive and of lower quality (as in: not as good/catchy). The best solution is probably to extract the semantic representation from raw/pcm form. (Incidentally, I'm working on this very problem. Anyone interested, feel free to contact me)

Generative music is already a thing, Jean-Michel Jarre's "En Attendant Cousteau" piece was released close to 30 years ago and Brian Eno coined the term a few years later: https://en.wikipedia.org/wiki/Generative_music

You can find a bunch of video on YouTube where Eno explains the whole process and how it works.

Idle thought: if musicians are being slowly replaced by algorithms, the future might not look like streaming individual tracks, but like apps that generate an endless amount of music locally. I think there'd be demand for such a thing from people who just want something on in the background. Possibly currently supplied by the four-hour or longer sets of electronic/classical music one sees on YouTube.

Reply on news.ycombinator.com