Reconhecimento automático de palavras

Abstract

Automatic word recognition is the translation of human speech into text, which has proved useful in man - computer communication. For this reason, several researches were developed in this area and consequently applications, such as virtual assistants, have arisen to make life easier for people. However, people who suffer from any dysphonia (hoarseness) can’t fully enjoy these applications, due to the distortions in their voice. This paper proposes the use of Convolutional Neural Networks (CNNs) to make the recognition of spoken words with this type of distortion. A database of 20 words with 28 samples was used, all voices were from different dysphonic people. Several topology were created for CNN, varying some hyperparameters of the network. All topologies were training and testing. For the test set, the topology with the highest accuracy obtained a result of 82,50%.