SongLLM

Concept demo Three prepared tracks, one of them instrumental. Readings by a model that heard the files, measurements by tools, one change applied for real.

Your song, read by an AI that listens.

Hear what a producer hears, pinned to the second. Choose one change, see exactly what it protects, and take the instruction to the tool you already use.

Jump to the fixes

0:00 / 0:00

Ask this track

Prepared answers. Nothing is sent.

    The reading

    Lyrics as prompted

    The lyric sheet is what was asked for, not proof of every word performed.

    
        
    
        
    
          Unedited model output
          
    
        
    
        
    
          Corrections and provenance
          
    
        
    
      
    
      
    
        
    Three changes, in order
    
        
    Each one says what was heard, what moves, what stays. Pick a destination and copy the instruction. One is applied for real so you can hear it.
    
        
      
      
      
    
      

    
      
      
      Across the three
      
      
      The same reading, side by side. This is what the layer does at catalog scale: every track gets a shape, a tempo, a key and a verdict before a human presses play.
      
      
      
    
      

    
      
      
      Why now
      
      
      Creation is solved. Suno, Udio and ElevenLabs will make you a song in a minute, and now they edit it too. Deciding what to change, and what to protect while you do, is not solved. That is the layer.
      
      
    
      

    
      
      
      Where this goes
      
      
      
        
      
          
      Inside the generator
      
          
      Every render gets a reading before the user hears it. Flat verse, buried hook, ending that ran out: pointed at, with a bounded change the tool can apply and a list of what it must not touch.
      
        
      
        
      
          
      At the label
      
          
      Ten thousand demos a month, each with a shape, a tempo, a key, a verdict and a timecoded note. A&R reads the ones the layer flagged, and hears the flag before pressing play.
      
        
      
        
      
          
      In the DAW
      
          
      A second opinion that listens to the bounce, not the session. Three changes in the language of the tool, each with what it protects, and an A/B at the moment it touched.
      
        
      
      
      
      
      
        
      SongLLM is a concept, this working demo and the domain. It is available to the team that should build it.
      
        
      
          Talk to us
          
        
      
      
      
    
      
  
      

  

  

  
    
      
      
      Upload is not live in this demo.
      
      
      Nothing you choose here is uploaded or analysed. The three tracks were generated with Suno and read once by an audio-native model that heard the files; the readings were then edited by a person and the numbers measured with tools. On a live build this button would:
      
      
        
        
      1. 1Listen. Your file goes to a music-understanding model, with tempo, structure and loudness measured separately by tools.
        2. 
        
      2. 2Point. Structure, tempo, key and the seconds that fail, pinned to the waveform, each labelled measured, estimated or heard.
        3. 
        
      3. 3Change. Up to three bounded changes, each with what it protects, as instructions for Suno Studio, ElevenLabs or your DAW. Then an A/B of your revision at the moment it touched.
        4. 
      
      
      
      
        
        Where this goes