TL;DR: Claude was able to generate audio with a single prompt. There was no need for multiple steps and a Python Script this time. Results sounded like this:
It’s been just shy of 2 years since Composing instrumental music with LLM’s part I. We’ve made some serious progress in those 2 years.
As a recap, in part 1, we prompted ChatGPT to create a melody and a drum beat. The output was fed into a Python Script that produced a midi file — and in turn, an audio output.
In part 2, Claude ends up not only generating melodies and a drum beat, but we get a really nice UI to play the results directly in the browser.
Prompt: you’ve entered a contest to create the best song ever. Create a melody that can be downloaded
The UI is really cool — and not something I asked for. The melody is a little underwhelming — however it doesn’t appear to be plagiarized like last time. Let’s see if Claude can alter the mood of the melody.
Prompt: make the melody really spooky
The UI changes — without me asking for that. It is once again very cool. Claude succeeded in shifting the mood — it’s quite foreboding. Let’s try out a drum beat.
Prompt: create a 808 machine hip hop beat
The UI update is awesome and reminiscent of an old school beat maker app on flip phones. The beat itself is simple but not bad. Let’s try to layer a melody and rhythm.
Prompt: Create a 16 bar melancholy melody over an EDM rhythm. Make the audio downloadable
Claude stacked the melody and rhythm nicely and it sounds very decent!
Closing thoughts:
LLM’s have improved in their ability to create music. Claude was able to change genres and tweak the emotional feel of the music.
Was the music itself incredible — not really. But an LLM is essentially a statistical madlib spewing text like a contestant on Who’s Line is it Anyway. It’s pretty incredible that an LLM can produce music at all.
There are music-specific AI companies, like Udio and Suno. These companies are creating the audio equivalent of AI-generated stock photos. They are also getting sued by the major music labels by training their models on copyrighted music.