Aprendizado profundo para pós-edição de saídas de sistemas de reconhecimento automático de fala

Abstract

Through the use of Artificial Neural Networks it is possible to perform voice recognition ofautomatically, allowing the conversion of speeches into text. This artifact is present on various devices, such as cell phones, smartphones and personal assistants, in addition to being widely used. Artificial Neural Networks have a number of nodes that are organized into layers, which are divided into input layer, hidden layer and output layer. These nodes represent neurons and intensify information processing. In this work, the construction of an artificial neural network that had with the aim of improving the output of an automatic recognition system of voice Applied in audios, which resulted as output a database containing phrases of Brazilian Portuguese. To fulfill the proposed objective Current Neural Networks of the Long Short-Term Memory type were used, as they present as a differential the have memory, that is, to carry, during the training process, information past at previous times and make use of this information to determine new values. Experiments were performed using databases that came from systems of automatic voice recognition in order to improve the network so that it can fulfill the purpose of being able tocorrect the sentences in the correct way. Then the construction of the final network was carriedout, changing the methodology used until then and obtaining better results.