An efficient noise-robust automatic speech recognition system using artificial neural networks

Abstract

The Human Auditory System for speech recognition is highly robust against background noise compared to state-of-the-art Automatic Speech Recognition (ASR) systems. One of the best ways to add robustness to a speech recognition system is to have a compressed and highly robust feature set. In this paper, we present a novel approach for feature compression which makes the proposed noise-robust ASR system simple and very efficient. The other popular approach for feature compression is K-means which is complex and time-consuming. The experiments were performed on the proposed noise-robust ASR system for recognizing 65 different words with very low Signal to Noise Ratio (SNR) of −5 dB. The Mel Frequency Cepstrum Coefficients (MFCCs) were used as features for speech recognition. Back propagation Artificial Neural Network (ANN) was used to design the ASR system. Specialized versions of feedforward neural networks and training algorithms were tested on the speech recognition system and results are presented for each type. Experimental results show that the recognition accuracy of the proposed noise-robust ASR system is very high even with such a low SNR.