gpt-3 MIDI Generator
This is a simple tool that uses gpt-3 to generate MIDI music notes and then renders those notes to a monotone midi.
This tool is an experiment to see if Large language models can pick up some music theory and generate music based on note patterns. It is NOT a replacement for a real music generator and there are definately more capable ways to generate music. Check out Magenta's Music Generation for a more robust music generator. Also OpenAI's Jukebox is a more robust music generator.
Nevertheless, this is a fun experiment and I hope you find it useful.
Listen to some samples here:
https://twitter.com/originalmaderix/status/1534951205339660290?s=20&t=B0dcYkyHiZ1rXMr1MG04dA
Requirements
- Python 3.6+
- pygame
- wave
- openai api
- numpy
Installation
To install, simply clone this repository and install the requirements:
git clone https://github.com/maderix/gpt-3-midi-generator.git
cd gpt-3-midi-generator
pip install -r requirements.txt
Usage
To use, simply run python gpt3_midi_generator.py. By default, this will generate 100 notes for the Star Wars theme. The generated notes will be saved to notes folder, and a wav file will be generated and saved to generated folder.
You can customize the behavior with the following arguments:
--model
The model to use for generation.Currently, the following models are supported: 'text-davinci-002' 'text-curie-001' 'text-babbage-001' 'text-ada-001'
--prompt
The prompt to use for gpt-3. This should be a string of space-separated phrase like 'Star Wars theme' or 'The original spiderman theme'.
--length
The length of the generated MIDI in notes.
--temperature
The temperature to use for gpt-3. This should be a float between 0 and 1, where 0 is the most deterministic and 1 is the most random.
--play_song
Whether or not to play the generated song. This requires pygame to be installed.
Mixer settings
--channels
Default is 2.
--sample_rate
Default is 44100.
--sample_width
Default is 2.
More settings to be added in future
OpenAI API
To use this tool, you need to sign up for an OpenAI API key. You can do so [here](https://openai.com/api/).
The request made by this tool WILL INCUR BILLING TO YOUR OPENAI ACCOUNT! Please measure usage accordingly.
Alternatively, you can use a smaller model for generation (curie, babbage, etc.) by using the --model flag.
In my experience, the curie is almost as good as davinci but incurs significantly less cost.
Issues
Please report any issues by raising a bug request on this repo.
License
MIT License