4 ms·

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

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.

exDM69 7y ago
> 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.
svantana 7y ago
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)

exDM69 7y ago

> 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.

svantana 7y ago

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)