Performance evaluation of front-end algorithms for robust speech recognition

Conventional speech feature extraction front-end algorithms (LPCC, PLP and MFCC) suffer severe performance degradation in noisy environment, especially when there is a noise level mismatch between the training and testing environments. Two more recently developed algorithms, namely Gammatone Cepstral Coefficients (GTCC) and Zero-crossings with Peak Amplitude (ZCPA), are claimed to have better performance than the conventional algorithms. To verify the claim, HMM-based speaker-independent continuous speech recognition experiments are conducted using TIMIT database. In these experiments, training data is kept in clean condition while various levels of white Gaussian noise are added to the testing data. Results suggest that GTCC outperforms PLP, which is the best amongst the three conventional algorithms.. by 1.6% in 0dB SNR to 4.4% in 20dB SNR. While ZCPA does not perform well in clean conditions, it performs better than PLP by 1. 5% in 20dB SNR to 4. 1 % in 0dB SNR. However, it has much higher computational complexity than all other evaluated algorithms.