This repository contains the works and laboratories of Speech and Language Processing I do in the Master's degree in Advanced Telecommunication Technologies (MATT), at Universitat Politècnica de Catalunya.
This report shows the study and development of differentmodels for language identification.We are comparing theperformance of a RNN baseline system with other models forthe same task. We will study modifications of the basic RNNmodel, as well as other DNN architectures
This report shows the study of transforming word vectors or word embeddings. So we create, analyze, and evaluate these word vectors. They are numerical representations about contextualized words into a learned vocabulary. Specifically we are trying to get the Catalan Word Vectors by training the following model.\ To do that task we are using a database created from Wikipedia articles.
This report shows the study and development of different models for non-causal language modeling. We are comparing the performance of a simple Feed Forward Neural Network with a Transformer Layer. Therefore we are increasing the number of Transformer Layers of the system and also using Multihead Attention. Finally we have applyed a normalization of Input/Output Shared Embeddings.