spoofify

Genrenator is a free public API that generates music genres that don't exist... but maybe should!

Examples:

• rhythm and euphonium chill
• punk didgeridoo
• czech pianostep

Don't these sound just like your next favorite genre that you would love to show up in your favorite music service?

Except... How do you find a performer who really represents your new favorite genre, so that you know where to start exploring it?

What are their top tracks? (You'll get to their back catalog and deep cuts eventually, of course, but I find that it helps to start with the more popular songs.)

And who even is in this musical collective?

I have no idea! They're not a real band (yet)!

But I know an excellent kind of tool that would gladly tell me things that don't exist.

So let's use it to find your next musical obsession!

Run the project

Prerequisites:

• Rye

You may want to run rye self update if you haven't recently.

Run locally

Create a .dev.env file with

SPOOFIFY_LLAMA_URL=<url_of_llama_instance>`

• rye sync
• rye run devserver

(You can also set the environment variable manually and use rye run spoofify)

You can either run your own local Ollama (with llama3.1:8b-instruct-q8_0, currently hardcoded) or ask @anna-hope for a URL to her own instance (if you're a Recurser or another friend, you should know where to find her!)

Run in production

(This is important for any serious application.)

• Create .prod.env with SPOOFIFY_LLAMA_URL that points to the hyper-optimized production instance of your LLM
• rye run prodserver

Run tests

You can run the extensive test suite with

rye test

(Uses pytest)

Format

rye fmt

(Uses Black)

Lint

rye lint

(Uses Ruff)

Dockerize

Since we want to keep the ability to install the package (so we can import it and use with hypercorn to serve in production, or with pytest), we need to follow these instructions from Rye's documentation

Concretely, this currently requires manually building the package before you can run docker build (not optimal, I know):

rye build --wheel --clean
docker build . --tag spoofify
docker run -p 8000:8000 -e SPOOFIFY_LLAMA_URL=<llama_url> spoofify

(The above may need more work depending on how ollama is being served)

LLM notes

I chose llama3.1:8b-instruct-q8_0 because I found it to have a good balance between being fast enough to get a response, understanding the prompt well enough, and good enough at returning a response in the expected format.

This model requires about 10GB of VRAM (on my M2 Max).

Because it is an LLM, it might occasionally generate responses that satisfy the expected format, but are otherwise strange or otherwise not ideal. Or sometimes it might completely fail to do what we want.

Example outputs

Wouldn't you want to scream your lungs out to these anthems of a forever-lost generation?

{
      "band_members": [
            "Vinnie Grits",
            "Lizzy Misanthrope",
            "Jesse Riffington",
            "Mike Dirtdrinker",
            "Sammy Scourge"
      ],
      "band_name": "Bleakstar",
      "genre": "grungetimism",
      "top_songs": [
            "Slumming It",
            "Gutter Revival",
            "Fuel for the Fire",
            "Blackout in a Box",
            "Riot's Revenge"
      ]
}

... or let your mind drift off into the pleasant musical abyss to this?

{

      "band_members": [
            "Lyra Lumen",
            "Kai Rhythm",
            "Aria Waves",
            "Nova Synth",
            "Caelum Beat"
      ],
      "band_name": "Echoflux",
      "genre": "neurochillhaus",
      "top_songs": [
            "Lost in the Haze",
            "Fractured Dreams",
            "Neuromantic",
            "Synthetic Skies",
            "Cosmic Drift"
      ]

}

or, I don't know, maybe you're in the mood to feel all the feels with these as your soundtrack?

{

      "band_members": [
            "Maxwell Wells",
            "Ava Morales",
            "Liam Flynn",
            "Ethan Patel",
            "Julia Knight"
      ],
      "band_name": "Echo Fade",
      "genre": "heavy singer-songwriter",
      "top_songs": [
            "Whiskey in the Rain",
            "Ghosts of Summer",
            "Fading Fast",
            "Lost and Found",
            "Burning Skies"
      ]

}

... or headbang to whatever this is

{

      "band_members": [
            "Vinnie 'The Bull' Blackwood",
            "Lars 'Lethal' Larsen",
            "Emily 'The Enforcer' Evans",
            "Jaxson 'Jolt' Jackson",
            "Maxwell 'Maelstrom' Martin"
      ],
      "band_name": "Riffwreck",
      "genre": "aggrohaus",
      "top_songs": [
            "Burning Inferno",
            "Cathedral of Chaos",
            "Echoes of Anarchy",
            "Hammerfall of Hell",
            "Terror in Tones"
      ]

}

(Honestly, I could keep adding these forever)

... sometimes the model decides to get a bit too creative:

{

      "band_members": [
            {
                  "instrument": "Lead Vocals, Synth",
                  "name": "Aurora Vex"
            },
            {
                  "instrument": "Drums, Percussion",
                  "name": "Kai Riven"
            },
            {
                  "instrument": "Bass Guitar",
                  "name": "Lena Lyrax"
            },
            {
                  "instrument": "Guitar, Effects",
                  "name": "Caspian Noir"
            }
      ],
      "band_name": "Echo Flux",
      "genre": "dance-techno whittle",
      "top_songs": [
            "Whittle Frenzy",
            "Fractured Rhythm",
            "Echo Chamber",
            "Techno Tectonic",
            "Lost in the Whirl"
      ]

}

Like, thank you for telling us who does what, llama, but now you've broken all of our clients' applications with this response.

(Maybe I should make it return the instruments, though.)

Additional information

This project was originally created as a way to demonstrate the project management capabilities of Rye versus other Python tooling. The code is not really the point, but I did try to make it do something fun.