Speech Recognition Using Artificial Neural Network

Abstract

This paper propose two approaches for speech recognition via supervised and unsupervised learning. Speech signals are non-stationary signals. Treating speech in computing domain falls under sequential learning task i.e. if we want to make a sense of current statement, we may need to go through the context in which it was spoken. Recurrent Neural Networks (RNN) is been used in speech recognition problems because of its powerful sequence modeling capacity. In this paper we have proposed Bi-directional Recurrent Neural Network with Long Short Term Memory model (LSTM), so that speech signal reconstruction can be done in a proper way without performance loss. For unsupervised learning, model is designed on the basis of Restricted Boltzmann Machine (RBM) which generates a reconstruction based output and helps in conversion of voice into text, each letter by letter.