Okay, so I saw a lot of music YouTubers fail miserably to make music with ChatGPT.
I luckily took some screenshots before OpenAi wiped their Chat History
ChatGPT is completely competent to make music, but you have to give it a few directions.
As a creative person, I talk to ChatGPT like it was a human, and not a machine.
So my first chats in a conversation usually looks like this:
“Can you help me make music?”
Where it replies, that it is ready to help me.
First, it tried to get to give me Guitar Tabs, but I’m a Piano Player.
So I asked it to give me Piano Chords.
Then I asked it to give me all chords in a chord key.
Which it managed completely fine.
In December I tried it to give me the notes for “Have Yourself a Merry Christmas”
However, it gave me notes in a key that didn’t fit the song. After an Hour Q&A with it, it apologized and claimed the original version was different than the version I played.
The way I got it to admit it faults, was to type the notes as Letters along with octave position:
G4-A4-B4 etc…
It tried multiple times to generate the old notes, but when I finally gave it the key scale, it gave up trying to correct me.
Now, onto the screenshots. (I expected them to be below the post, and not inside it)
I haven’t looked at the screenshots in ages, so I don’t remember the order I took them in, in the conversation.
That being said, it’s a very good composer.
I think I got it to make a song, and I meant melody, but it gave me lyrics.
Fair enough, my bad.
But since I already had the lyrics, I asked if to give me chords to accompany the lyrics.
I haven’t tested to play them yet, but the chord order seems to be correc notation.
Then I asked it for a melody, and it gave a very bland melody with few note changes. Very monotone soundscape.
I asked it to make it more singable, and it did.
Another discovery which was very pleasing, was that it could give chords in Roman Numerals.
It also gave me chords in Dorian, Mixolydian, Aeolian modes.
Conclusion:
It’s very capable of creating music, you just have to be very clear of what output you want to receive.
It’s sort of like a Teenager asking mom where a certain bowl is.
Mom answers: In the cupboard.
The teenager didn’t specific which bowl he was looking for, so mom answers in a non-specific manner as the cupboard is the most likely location for a bowl according to her dataset / memory.
By asking: “where is the popcorn bowl?” Mom can say which cupboard it is in.
It’s the same way with music. You have to give it specific instructions that usually is visual information.
If you give it a 4/4 beat, which key, the characteristics of the melody, what kind of chord progression.
It can create beautiful notation, and very handy tables of music theory.