Show HN: Melogen – Generate MIDI melodies for free • Hacker News Board

Show HN: Melogen – Generate MIDI melodies for free

#1

Post by squirrelon » Tue, May 19, 2026, 7:00 PM UTC

Melogen generates short melodic phrases as .mid files, no rendered audio, just notes you drop into your DAW. Free to try, no signup required.

I built this because every "AI music" tool I'd tried produces finished audio, which is the opposite of what I want when I'm stuck on a hook. I wanted something that hands me a MIDI seed I can take into Ableton/FL/Logic and shape myself.

Under the hood: - Decoder-only transformer trained on MIDI - Outputs note sequences with timing and velocity - Tuned for short, usable phrases rather than full arrangements

It's still in beta and rough in places. I'd value blunt feedback on: - Whether the generations are actually musical, or just "notes that fit a key" - MIDI handling: timing, velocity, channel assignment - What's missing before it would slot into a real workflow

Happy to get into the model, dataset, or pipeline in the comments.

Show HN: Melogen – Generate MIDI melodies for free
melogen.ai

Re: Show HN: Melogen – Generate MIDI melodies for free

#2

Post by squirrelon » Tue, May 19, 2026, 7:39 PM UTC

Feel free to comment any enhancement suggestions or points you have.

Re: Show HN: Melogen – Generate MIDI melodies for free

#3

Post by Rochus » Tue, May 19, 2026, 9:21 PM UTC

squirrelon wrote: post #2

Feel free to comment any enhancement suggestions or points you have.

Cool. Generation of symbolic music using transformers is indeed a pretty neglected field. I assume you have to encode more "musical knowledge" into the embeddings than when you "just" compress waveforms. Can you provide information on your embeddings?

Re: Show HN: Melogen – Generate MIDI melodies for free

#4

Post by squirrelon » Tue, May 19, 2026, 11:16 PM UTC

Rochus wrote: post #3

squirrelon wrote: post #2

Feel free to comment any enhancement suggestions or points you have.

Cool. Generation of symbolic music using transformers is indeed a pretty neglected field. I assume you have to encode more "musical knowledge" into the embeddings than when you "just" compress waveforms. Can you provide information on your embeddings?

Hi, It is actually not using transformers, those would be too slow. It is using a combination of CNN's and linear layers. Correct, it uses embedings, not waveforms or spectrograms. The inputs are midis, some of which I made myself in FL Studio. The model creates a "latent representation" from each midi, I can then sample randomly from this latent space to get an original piece. The most important part is the preprocessing in my opinion.

Re: Show HN: Melogen – Generate MIDI melodies for free

#5

Post by Rochus » Wed, May 20, 2026, 12:32 AM UTC

squirrelon wrote: post #4

Rochus wrote: post #3

Earlier quoted context omitted.

Cool. Generation of symbolic music using transformers is indeed a pretty neglected field. I assume you have to encode more "musical knowledge" into the embeddings than when you "just" compress waveforms. Can you provide information on your embeddings?

Hi, It is actually not using transformers, those would be too slow. It is using a combination of CNN's and linear layers. Correct, it uses embedings, not waveforms or spectrograms. The inputs are midis, some of which I made myself in FL Studio. The model creates a "latent representation" from each midi, I can then sample randomly from this latent space to get an original piece. The most important part is the preproce…

That's fascinating. This sounds like a variational autoencoder. The embeddings, which from my humble point of view (as a trained musician) are a largely unexplored field not really supported by existing theory, are at the same time game-deciding. Have you found a good solution for this?

Re: Show HN: Melogen – Generate MIDI melodies for free

#6

Post by squirrelon » Wed, May 20, 2026, 9:05 AM UTC

Rochus wrote: post #5

squirrelon wrote: post #4

Earlier quoted context omitted.

Hi, It is actually not using transformers, those would be too slow. It is using a combination of CNN's and linear layers. Correct, it uses embedings, not waveforms or spectrograms. The inputs are midis, some of which I made myself in FL Studio. The model creates a "latent representation" from each midi, I can then sample randomly from this latent space to get an original piece. The most important part is the preproce…

That's fascinating. This sounds like a variational autoencoder. The embeddings, which from my humble point of view (as a trained musician) are a largely unexplored field not really supported by existing theory, are at the same time game-deciding. Have you found a good solution for this?

[flagged]