Robust feature extraction methods for speech recognition in noisy environments

This paper presents robust feature extraction techniques for isolated word recognition under noisy conditions. The proposed hybrid feature extraction techniques are Bark Frequency Cepstral Coefficients (BFCC) and Weighted Average Mel-Frequency Cepstral Coefficient (WMFCC). Both methods are tested in various noisy environments using a single Gaussian Hidden Markov Model (HMM) based isolated digit recognition system. The results clearly indicates that WMFCC performed well compared to Mel-Frequency Cepstral Coefficient (MFCC) in noisy environment using NOISEX-92 database.