
AI Music From the Bottom Up: Why MIDI Lab's Audio Is Not AI Generated
Two directions
There are two ways to make a song with AI, and they run in opposite directions.
Top down. A model trained on finished recordings generates a finished recording: one waveform with the drums, bass, chords and vocal already baked together. If you want stems, another model has to pull that waveform apart afterwards. To see the two directions side by side, compare a popular song maker with MIDI Lab.
Bottom up. The song is built the way producers build one. First the notes, part by part. Then a sound for each part: a synth patch, a drum kit. Then the audio, one part at a time. Nothing is mixed until you mix it. For tools that give you the notes, see song maker alternatives for producers.
MIDI Lab works bottom up.
What the AI does, and what it does not
At MIDI Lab the AI writes music as MIDI: which notes, when, how hard, on which lane. That is where it stops. The sound you hear comes from synthesizers and samples playing those notes, the same way a synth in your studio plays the MIDI you send it. Even when the AI picks or designs a synth patch, the sound itself is made by the synth.
When you download audio, each lane is rendered on its own from that MIDI. So the audio is synthesized, not generated. No AI model draws the waveform, and no model separates it into stems. The drum stem is the drum lane played through the kit. The bass stem is the bass lane played through its patch.
Why bottom up suits producers
Every part is really separate. A stem split from a finished mix can carry a trace of the other parts. A lane rendered on its own holds that part and nothing else.
You can still change it. The MIDI comes with the audio. Rewrite a bassline, change a chord, swap a patch, then render again. A generated waveform cannot be edited note by note. Already have a .mid file? The free MIDI to MP3 converter turns it into audio in your browser.
It ends up yours. The point is a starting point you can take apart and rebuild: your sounds, your edits, your arrangement, your mix. Here is how MIDI Lab compares with another song maker on downloads and ownership.
What streaming services are doing about AI audio
Streaming platforms are responding to the flood of fully generated tracks, and the details matter if you release music.
Deezer says its detection tool looks for signatures that generative AI models leave in the audio, and that it currently focuses on music created entirely by generative AI platforms. Tracks it detects are tagged and left out of algorithmic recommendations and editorial playlists. Read Deezer on AI labelling.
Spotify supports an industry standard for AI disclosures in song credits, developed through DDEX, filters spam and mass uploads, and does not allow unauthorized voice clones. It has said that disclosure is not about punishing artists who use AI responsibly. Read Spotify on AI protections.
Distributors set their own rules too, and they change often. Some ask you to disclose any AI involvement, including in the songwriting.
Releasing music you started in MIDI Lab
Be straightforward about it. The audio from MIDI Lab is synthesized from MIDI, not generated by an AI audio model. The notes, though, start with AI. So:
- Read your distributor's current rules on AI before you upload, and disclose AI assistance where they ask for it.
- Make it yours. Rewrite parts, replace sounds with your own, record over it, arrange and mix it. The more of the song is your work, the more it is your song.
- Keep your project files. Your MIDI, your edits and your DAW session show how the track was made.
Try it
Generate a song, then use the Download button on its card and choose Audio to get every lane as its own file. For more on rendered stems and split stems, read Stems From MIDI vs Stems Split From Audio.