An AI that listens to your song and gives you notes
A live concept demo at songllm.com. An audio-native model hears a track and returns a producer's reading pinned to the waveform, with three suggested changes and the instruction for the tool that would make them.
Visit songllm.com
The purpose
Music generators and editors are plentiful. What a songwriter lacks is a critic: something that hears the track, says where it goes flat and why, and hands over a change they can act on without a producer in the room.
Heard by a model, measured by tools, chosen by you
A track goes through a model reading and tool measurements. The listener chooses one change and takes the instruction to the tool they already use.
Example route: a listener picks a change and hears it applied.
The complete process is shown.
Arrow keys step through the route. Home returns to the full view.
One example route. All alternatives stay visible.
Read the process and handoffs
- Choose a track
Listener · Track list.
- Audio file: Read the audio.
- Audio file: Measure the track.
- Read the audio
Model · Audio-native model. Output is corrected against the lyric sheet and file length.
- Model reading: Label each number.
- Measure the track
Tools · Tempo and loudness.
- Measurements: Label each number.
- Label each number
Fred · Measured, model, disputed.
- Labelled facts: Show the reading.
- Show the reading
SongLLM · Waveform and notes.
- Three changes: Pick one change.
- Pick one change
Listener · Heard, changes, stays.
- Applied change: Hear it applied.
- Copied instruction: Copy to a tool.
- Hear it applied
Listener · A/B player.
- Copy to a tool
Listener · Suno, ElevenLabs, DAW.
My part
I set the critic-not-editor frame, ran the model readings and the measurement scripts, wrote the corrections ledger, applied the audio change, and built and shipped the page.
In practice
Live and in a first round of tester feedback. The upload flow is scoped for now: the page explains what a live build would run instead of pretending to run it.
The readings come from a model and are labelled as such. Key detection is marked disputed where the tools disagree. Nothing on the page claims more than it can show.
Want to know more?
This one is public. The source stays private, but the product is live and you can try it. Get in touch if you would like to discuss how it was built.
Ask about this project