Abstract:

In this paper, the authors describe a lossless audio signal compression solution using a variational autoencoder. The convolutional neural network layers in the encoder convert the graphical form of the audio signal spectrogram to latent space features. Experiments using the PyTorch library for neural networks with signal processing using the Libros library are discussed. In the next steps of the study, the authors want to adjust the convolution filters dynamically depending on the spectrogram resolution.