AP Research

Synthetic generative techniques like Generative Adversarial Networks(GANs) and other machine learning models have been used to create new data from old data by learning the latent representation and structure of the training data. This technology is relatively new, and as a result, more research needs to be conducted regarding its potential applications across the board. However, the results are promising as GANs and other generative networks can be used across disciplines ranging from health care, financial technology, and other business sectors. In addition, this approach has been used in producing AI-generated art and music. But how well do these different models compare against each other when trying to make classical, rock, and pop music? This project aims to compare three different machine learning models for the purpose of synthetic music generation. The main model architectures I plan on using are Convolutional Neural Networks, Generative Adversarial Networks, and Variational Autoencoders, as they have been used frequently for synthetic generation tasks. I decided to try and gear all the training data towards mainly instrumental music as the lyrical aspect of many modern pop and rock songs add an additional layer of complexity to this project. I plan to compare the results of the generated audio by comparing the waveforms between artificial and real music and a survey of 50 people asking them which model they think performed the best in generating music from a specific genre. I will also rank these models in computational efficiency relative to the accuracy and level of artificial music they generate.