This project explores storing an audio signal in the parameters of a neural network. The model takes a position in the signal and predicts the corresponding audio value.
The workflow fits the model, reconstructs the audio, and measures the result. Parameter sweeps compare different model sizes and architectures.
The practical tradeoff is between the size of the model, the time needed to fit it, and the quality of the reconstructed sound. The project is an experiment with that representation. It does not yet establish an advantage over conventional audio codecs.