Super-resolution of Audio Files Using Feed-forward Neural Networks for Music Storage and Transfer

In this report, a system for reducing the file size of an audio signal, and then performing super-resolution on the resultant signal to estimate the original, is proposed and designed. This design takes influence from the principles of audio sampling, as well as super-resolution systems designed for visual media, and is split into an encoder and a decoder. The encoder successfully reduces the file size of the audio file by a significant amount. The super-resolution-based decoder can also successfully generate a matching high-frequency audio track that can be combined with the encoded lossy audio in order to estimate the original audio with a reasonable degree of accuracy. While a number of improvements to the system can be made in the future, it shows great promise, as it accomplishes the goals it was designed to meet.