I'm working on a framework that connects multiple publicly available AIs (GPT-4, Suno Chirp, several Sentence Transformer models, etc.) in an effort to create a fully autonomous singer/songwriter. The idea is that a script can run, and when it's done, the result is an original song, complete with title, genre, lyrics, and full audio.

Currently, it's about 90% automated. It does all the topic brainstorming, genre choice, and lyric writing on its own, including building up a memory bank of imagined (but self-consistent) personal experiences it can draw upon for future songwriting inspiration (using a handful of semantic search techniques). And then Suno Chirp generates the audio.

The only part that's not automated right now is the connection between the lyrics/genre and the audio. I'm currently manually copying the generated lyrics into Suno Chirp and curating the best generations from each 30-second chunk to produce the final song.

I'd like this to be 100% automated, which brings me to the question: does anyone know of an API or library (Python preferred, but any language will do) that can be passed a short audio clip, then grade it on "quality"?

Equivalently in this case, a classifier that can grade the audio on "likelihood of being AI generated" would also work, since the same qualities that Chirp produces that "sound bad" would likely raise such a classifier's response.

I've tried searching everywhere, and I've only found one research paper that used human biofeedback to determine how good a song is (obviously not useful here), and one service that "detects AI music" by just keeping a human-submitted database of known AI clips (also not useful).

As soon as I have a way to automatically determine a quality score for a 30-40 second music clip, I can fully automate the system, which would be a major win for me.

So does anyone know of a way to do this?