Do androids dream of electric beats?
I’m sure I could do this on my own with normal ML techniques and training my own NLP model from scratch, but I want to see if GPT-3 can do it first.
The idea is that because MIDI files are well-structured and have a lot of descriptive metadata in it, it should theoretically be rather straight forward to use C to read the binary and convert it to a human-readable text format with a metadata header and then lines of descriptive text for each note. In that case, it is a NLP problem to get an AI to understand the text file, and then output one of its own similar to the original with metadata and a few notes as a prompt. It should then theoretically be able to produce music based on the topic, and that text file is then converted back into MIDI then played.
How would I go about doing this using GPT-3 or is it better for me to do this on my own with pure Python to train the models, without the help of any API?
