Detecção de fala e comandos baseada em redes neurais convolucionais

Abstract

The present work consists in the development of and artificially intelligent system, based on artificial neural networks, destined to the real-time classification of audio, aiming to identify the presence of speech and an activation command in a microphone input. The scope of the application here proposed comprises the labeled audio samples acquisition for supervised learning, the development of peripheral algorithms for data management with the use of melscale spectrograms, as well as the construction, training and validation of convolutional and convolutional-recurrent neural network models. Validating the models by its accuracy obtained during the training process against the validation dataset, it is possible to verify that both models are able to learn, and the purely convolutional model excels in all training cases.