AI and Future of Music Production

[CapMan]

I taught myself to play bass by listening to Steve Harris's bass lines in Iron Maiden. Was I stealing his work/content by listening and learning how he played and how to play in a similar way?

Did you learn to play drums without ever hearing how another human played drums?

It's interesting to me whether your experience of learning and AI's experience of learning have some kind of equivalence , which is what think what you are saying. Personally I don't feel they are the same, but what's your take ?

Of course humans listen to other musicians and learn from them. I imagine that you and I did this for similar reasons - it helped us play better, gave us purpose, we found it fulfilling, it was fun, it meant we had the skills to play in groups with other humans and interact with them.

AI isn't doing any of those things (maybe it will in the future) so don't think there is any equivalence between how we 'steal' from others and how/why AI does it I may be wrong though, Interested in what others think .

It's great to be able to have this conversation and share perspectives in a thoughtful way. Thanks

 

[Brian Hall]

It's interesting to me whether your experience of learning and AI's experience of learning have some kind of equivalence , which is what think what you are saying. Personally I don't feel they are the same, but what's your take ?

Of course humans listen to other musicians and learn from them. I imagine that you and I did this for similar reasons - it helped us play better, gave us purpose, we found it fulfilling, it was fun, it meant we had the skills to play in groups with other humans and interact with them.

AI isn't doing any of those things (maybe it will in the future) so don't think there is any equivalence between how we 'steal' from others and how/why AI does it I may be wrong though, Interested in what others think .

It's great to be able to have this conversation and share perspectives in a thoughtful way. Thanks

I think most people who learn to play do it because they find it enjoyable. As with most things, we learn new things based on what others have done or discovered. AI doesn't learn because it wants to play, it does as instructed (so far ). AI music models are designed to be a tool for humans to use.

The AI music tools can be used in a very simplistic way (push a button and get a song in some style) or in a way more similar to working with hired musicians. I prefer the later approach.

My point in the previous post was that AI "listens" to music samples to "learn" what a guitar sounds like. It also learns about how beats per minute work and how different key signatures sound. I don't see that as stealing any more than I was stealing from Steve Harris when I learned his bass lines by listening. It's pattern recognition and learning, just at a much larger scale.

These tools can already listen to a reference track of me playing a guitar, bass, keyboard or humming or badly singing a melody and use that to guide the model as it generates the other audio parts.

I don't think it will be long, maybe a year or so until I will be able to set up a jam session (or live performance) with me playing an instrument and have the AI playing along with me (following my lead and instructions) in real time using designated instruments.

 

[CapMan]

@Brian Hall Good to get your perspectives and learnings as always.

 

[Brian Hall]

New song. I usually get an idea for lyrics first and then generate an image to go with it. This time, the image came first and got me thinking about the lyrics. The music and vocals were generated with Ace-Step Studio using a 1.5 xl-turbo merge model on my pc.

 

[Brian Hall]

I noticed an audio glitch in the original version of this song. I fixed the glitch (A sharp click or pop) with Audacity's repair tool and uploaded the fixed version. Nothing else has changed.

 

[HalSF]

The New York Times has splashed out with a feature on DIY remixing using AI (gift link):

An A.I. Music F.A.Q.: Can I Remix Madonna? Is This All Legal?

Advances in A.I. are making it possible to create all kinds of music from scratch, but they also raise questions about what is legal and who will actually listen.

www.nytimes.com

 

[KSTR]

New song. I usually get an idea for lyrics first and then generate an image to go with it. This time, the image came first and got me thinking about the lyrics. The music and vocals were generated with Ace-Step Studio using a 1.5 xl-turbo merge model on my pc.

While is is absolutely fascinating what one can do at home right now with these tools, I think this track also showcases some limits wrt sound quality.

To me, the track sounds OK when played back on a laptop while doing kitchen work, it would also be OK as background music in a snack bar, or as a demo tape... but once you listen a bit more carefully on a good system things fall apart quickly, notably when one has a musical background and plays a number of instruments.

It may also have to do with YouTube's compression but what I note is a strong annoying constant background "garbling" noise, and the acoustic and electric guitars sound very artificial, sort of partly polished but lifeless and quirky, weird attacks, no fretting/sliding noise, etc. The vocals have that auto-tune kind of sound signature and again are a bit lifeless in timbre, trying to sound emotional but not really getting it across. Drums are OK except that tom fill. Keys and bass sound OK.

Of course this is a problematic track with that guitars intro which exposes their weird sound much more than it would in a more dense arrangement.

So overall, quite a success in relative terms but would I want to listen to that for recreational enjoyment? I don't think so.

 

[Brian Hall]

While is is absolutely fascinating what one can do at home right now with these tools, I think this track also showcases some limits wrt sound quality.

To me, the track sounds OK when played back on a laptop while doing kitchen work, it would also be OK as background music in a snack bar, or as a demo tape... but once you listen a bit more carefully on a good system things fall apart quickly, notably when one has a musical background and plays a number of instruments.

It may also have to do with YouTube's compression but what I note is a strong annoying constant background "garbling" noise, and the acoustic and electric guitars sound very artificial, sort of partly polished but lifeless and quirky, weird attacks, no fretting/sliding noise, etc. The vocals have that auto-tune kind of sound signature and again are a bit lifeless in timbre, trying to sound emotional but not really getting it across. Drums are OK except that tom fill. Keys and bass sound OK.

Of course this is a problematic track with that guitars intro which exposes their weird sound much more than it would in a more dense arrangement.

So overall, quite a success in relative terms but would I want to listen to that for recreational enjoyment? I don't think so.

It may partially be youtube. Also, the first version I uploaded had an audio glitch. I fixed that and uploaded another version of it. I just listened to the original flac again, (the one after I fixed the audio glitch pop / click). It sounds pretty clean to me on my Klipsch Heresy IVs. I forgot to mention in the description of the song, but the bass track is actually me playing my 1983 Fender Jazz Bass Special. The guitar is supposed to sound like a 6 string classical acoustic guitar, not a regular acoustic or electric.

When I finished the song, I thought it was just ok. After waiting a few days and listening again, I like it a lot more. One of my more artistic friends texted me and said it made him "smile and cry at the same time" so mission accomplished.

Now I'm working on new lyrics for a new song based on a ghost image I generated and a ghost ballet video I had grok generate:

 

[KSTR]

I forgot to mention in the description of the song, but the bass track is actually me playing my 1983 Fender Jazz Bass Special.

Great ;-)

The guitar is supposed to sound like a 6 string classical acoustic guitar, not a regular acoustic or electric.

To me, acoustic means classical Spanish nylon string, whereas your regular steel string acoustic would go under Western or Dreadnought moniker.
The type of guitar was immediately obvious, as was the quite Fender'ish bright tone of the electric.

 

[DanielT]

It’s nuts to say, but I think they’re getting better:

 

[Brian Hall]

To me, acoustic means classical Spanish nylon string, whereas your regular steel string acoustic would go under Western or Dreadnought moniker.
The type of guitar was immediately obvious, as was the quite Fender'ish bright tone of the electric.

I think of a spanish guitar as distinct from an acoustic guitar. The nylon strings vs the metal strings give a very different sound and feel.

 

[KSTR]

Acoustic guitar is the general class, denoting the operational principle. Embodiements would be:
- Classical/Spanish/Flamenco guitar with Nylon or, originally, gut strings
- same, but with steel strings, popular in Latin America
- Western/Dreadnought, popular in Northern America, steel strings
- Gipsy (Selmer/Maccaferri) guitar, steel strings
- Archtops with f-holes, as used in early Jazz etc., steel strings
- whatever I forgot (including the many subvariants of the above).
All sound different.

 

[DeadBeatAnon]

I do like this AI voice morphing used to create a new 'McCartney' song from Beatles session fragments of ideas. The first half describes the process, the second half is the song. Some lines sound more like him than others, but getting closer...

It sounds more like a McCartney & Wings song. Even though the song originally pops up in a '69 Beatles session, there's a reason it didn't land on a Beatles album. Think about it--on what late Beatles album would "Pillow for Your Head" find a home. It doesn't fit anywhere on Let It Be or Abbey Road. Even if you rewind to '68, it doesn't fit on the The Beatles (White) album, which has all kinds of experimental tracks. Reaching back into the Beatles earlier catalog, it doesn't fit the transitional period of Revolver or Rubber Soul. Maybe 1965 or earlier (Help! or Beatles for Sale). By 1969 the only home for this song would've been on the Yellow Submarine soundtrack, an ad-hoc collection of songs--widely considered to be the band's worst album.

 

[Brian Hall]

I finished a new song. I've been working on the lyrics off and on for several days. The song is called "One Last Dance" and is based on an image I generated of a ghost woman. I put full details in the description. I used my lyrics and generated the music and vocals in Ace-Step Studio running on my pc using a merged model of Ace-Step 1.5 turbo and Ace-Step 1.5 xl-sft. It took 68 generations to get this version over about 8 hours (with breaks).

 

[DeadBeatAnon]

Check the 1:36 mark--I think the subtitle should be "motes hang still", not "modes hang still". Also check the 1:06 subtitle, maybe "all my memories sleep" instead of "all my memory sleep" (plural vs singular). Then at the 3:37 mark, try "No audience waits in the empty rows" instead of "No audience waits in the empty rose". There may be more subtitle issues. Are you using a speech-to-text model? If so, the AI model is having issues with homophones (rows/rose) and similar-sounding phonemes (motes/modes).

Nicely done, although I was jarred by the antiquarian image paired with a modern rock track. Have you considered using the same song with a different arrangement, perhaps just acoustic guitar, piano, and violin. It might be an interesting experiment. Good luck.

 

[Brian Hall]

Check the 1:36 mark--I think the subtitle should be "motes hang still", not "modes hang still". Also check the 1:06 subtitle, maybe "all my memories sleep" instead of "all my memory sleep" (plural vs singular). Then at the 3:37 mark, try "No audience waits in the empty rows" instead of "No audience waits in the empty rose". There may be more subtitle issues. Are you using a speech-to-text model? If so, the AI model is having issues with homophones (rows/rose) and similar-sounding phonemes (motes/modes).

Nicely done, although I was jarred by the antiquarian image paired with a modern rock track. Have you considered using the same song with a different arrangement, perhaps just acoustic guitar, piano, and violin. It might be an interesting experiment. Good luck.

I didn't add the subtitles. Youtube creates them automatically. Your interpretation of the words is correct.

Edited to add: The very first generation was acoustic guitar, bass and vocals which can drift towards a country music style which wasn't what I was going for.

 

[Brian Hall]

Based on the message by @DeadBeatAnon I decided to generate a different version of the same song. This version uses a different style, bpm, instruments and model. I used the Ace-Step 1.5 XL-Turbo model instead of the merged model I used in the original. I didn't change my lyrics.

 

[DeadBeatAnon]

Beautifully done. Enjoyed listening to that at 2 am with my headphones on. And that still image is quite haunting. The young woman reminds me of Nicole Kidman in The Others.

 

[Brian Hall]

Beautifully done. Enjoyed listening to that at 2 am with my headphones on. And that still image is quite haunting. The young woman reminds me of Nicole Kidman in The Others.

Thank you.

 

[Brian Hall]

I'm trying out a new locally installed AI Music model called YuE2 that is much more advanced than Ace-Step 1.5. It has the capability of accepting detailed music instructions in something called ABC format. YuE2 uses a limited version of the full ABC format, but gives much more options than Ace-Step. ABC is a text format where letters are notes, numbers are lengths, and a short header sets key, meter, and tempo. A lot of good and interesting things to learn that will take a little while. There are utilities for converting ABC to sheet music and vice versa.

This is an example of what the ABC music text format looks like:

X:1
T:
M:2/4
L:1/16
Q:1/4=75
V: Vocal clef=treble name="Vocal Melody" snm="Vocal"
V: Ins clef=treble name="Ins Melody" snm="Inst."
K:Bb
% intro
V: Vocal
z4"Bb"z4|"Bb"z8|"Bb"z8|"Bb"z8|
V: Ins
Z|B,B,B,F2B,zF-|FB,zFDE2F|Dz2F2B,zF-|
V: Vocal
"Bb"z8|"Bb"z8|"Bb"z8|"Bb"z8|
V: Ins
FB,zFDE2F|D3F2z2F|z3FDE2F|D3Fz4|
V: Vocal
"Bb"z4Bd2d-|
V: Ins

It can take a reference track to follow, which Ace-Step can also do.