SoundNet-keras

The purpose of this notebook is to explain how SoundNet works (maths, code, and experiments). Soundnet was developed in 2016 in order to use the natural synchronization between pictures and sounds to learn an acoustic representation from a large number of unlabeled videos, it means to get important features of the sound which allow us to depict them. Other studies are focused on features such as spectrograms and MFCC, on another hand, Soundnet is focused on the natural sounds.

Deep-Learning

These lab sessions are based on the code given by the teacher staff. Our work is understand the principles and make experiments who compare differents models of deep learning.

ciGAN

Second version ciGan. Data augmentation using Generation Antagonist Networks (GANs).

audiosummary-pytorch

Audiosummary using soundnet-pytorch and FMRI analysis