DEEP RECURRENT NEURAL NETWORK BASED AUDIO SPEECH RECOGNITION SYSTEM

Abstract

Speech recognition system has become an integral part of how the computer technologies can be used to influence and improve the human work and activities. These are being used extensively, from personal assistants to self-driving car Human Computer Interfaces (HCIs) and other industries as well. While the most common approach to speech recognition system building is using the Hidden Markov Models (HMMs). The HMM models assume a specific structure of the data and unable to capture temporal dependencies. This paper however presents a unique approach for isolated word recognition based on deep learning models using Recurrent Neural Networks (RNNs) particularly, which can perform end to end speech recognition without any assumption of structure in data using Bidirectional LSTM (BiLSTM). The network proposed in the paper can learn both features in the data and capture temporal dependencies.