Aplicação de redes neurais recorrentes no reconhecimento automático da fala em ambientes com ruídos

Abstract

Many learning tasks require dealing with sequential data, such as text translators, music generators, and more. Deep Neural Networks have shown promising results in automatic speech recognition, where one of the main challenges is voice recognition signals in the presence of noise. In this manuscript, we combine two known deep learning architectures, Convolutional Neural Networks for acoustic modeling, and a recurrent architecture with Classification Temporal Conexionist for sequential modeling. Recurrent Neural Networks (RNN) are models that capture sequence dynamics through a topology that contains cycles, unlike acyclic neural networks or feedforward networks. The RNN studied in this work is a particular case of a deep learning network that, unlike its shallow correlates, it is able to retain a state that can represent information from an arbitrarily long context window. The experimental results showed that the proposed architecture achieved superior performance when compared to Hidden Markov Model in tests carried out on the same databases.