Robust front-end and back-end processing for feature extraction for Hindi speech recognition

In this paper, we present a robust feature extraction algorithm based on auditory periphery model for Speech Recognition. At the front-end, a normalized filter bank based on Gammatone filtering is applied to the speech spectra followed by a power law non-linearity. Experiments show that the proposed features named as EFCCs (ERB scale cepstral coefficients) outperform MFCCs in noisy environments without losing performance in clean environment as well. To improve the performance further, a feature enhancement algorithm named as MVA (Mean subtraction, Variance Normalization and ARMA filtering) is applied at the back-end. A 238-word continuous speech recognition task was used to evaluate the proposed feature extraction algorithm. Tests conducted in presence of additive white Gaussian noise at different Signal to Noise ratios (SNR) reveal that EFCCs performed 10.5% (on average) better than MFCCs. Performance improvement in case of EFCCs-MVA was observed to be 16.1%.