A statistical analysis on the impact of noise on MFCC features for speech recognition

Noise is omnipresent in almost all acoustical environments. The investigation presents here seeks to quantify the impact of noise on mel-frequency cepstral coefficients (MFCC) of speech signal. MFCC is one of the most commonly used features for speech recognition systems. However, it has been observed that performance of MFCC based system degrades drastically with changing noise levels and noise types. In the present study, different noise types at different levels have been added to the clean speech signal and the changes in statistical distribution pattern of the signal has been investigated. Further, performance of two commonly used noise normalization techniques Cepstral Mean and Variance Normalization (CMVN) and Spectral Subtraction (SS) have also been evaluated.