Speech Analysis for Automatic Identification of Mild Cognitive Impairment through Autoencoders

Nowadays, cognitive decline is unfortunately a non-curable disease which, due to an alteration in brain function, causes the progressive decline of memory, thought and reasoning abilities, so much so that in its most severe state patients reach the complete loss of autonomy. Being able to identify the first signs of cognitive decline in a "pre-symptomatic" phase certainly becomes fundamental in trying to respond significantly to the disease. To succeed in this, the researchers focused on one of the most evolved abilities of the human mind: the language. In this thesis we proposed a method to classify audio files to detect subjects suffering from cognitive decline. In particular we used the Autoencoder method that is a type of artificial neural network used to learn efficient data codings in an unsupervised manner. The specific use of the autoencoder is to use a feedforward approach to reconstitute an output from an input. In our case we have used the software auDeep that is software for unsupervised feature learning with deep neural networks (DNNs), which trains an autoencoder for the extraction of features from spectrograms and their classification. In addition, the SpecAugment method was used to increase the number of data to be analyzed. This method involves cutting some frequency and time bands into the spectrograms and using them as new data.