Recognition of Isolated Words Using Feedforward Neural Networks

In this article we present a speech recognition system for isolated words based on neural networks and explore its behavior. We applied Feedforward artificial neural networks of different complexities and various sequence length normalization techniques. The system uses well-known MFCC speech features widely deployed in variety of speech applications, together with cepstral mean normalizations to help the classifier to reach higher accuracy and robustness. Apart of that we suggested eligible scenarios to trim unequal speech sequences to fixed length training inputs for Feedforward networks. Finally, an extra parameter of the length of original utterances was explicitly added to a feature vector to test its importance for the recognition process using neural networks. All experiments and evaluations were accomplished on the professional database used to train recognition systems and compared to a standard approach based on continuous density Hidden Markov Models.