Hindi vowel classification using GFCC and formant analysis in sensor mismatch condition

In the presence of noise and sensor mismatch condition performance of a conventional automatic Hindi speech recognizer starts to degrade, while we human being are able to segregate, focus and recognize the target speech. In this paper, we have used auditory based feature extraction procedure Gammatone frequency cepstral coefficient (GFCC) for Hindi phoneme classification. To distinguish vowels from each other, we have analyzed frequency response curves of each vowel. Here we propose a new feature extraction technique by taking first three formant frequencies of each vowel along with their cepstral features to increase the phoneme classification performance in noisy condition. The classification performance achieved by the proposed features is compared with the standard MFCC and GFCC based features using a continuous density hidden Markov model (CDHMM) with a mixture of Gaussian distributions. To evaluate robustness of these features in noisy environment, the NOISEX database is used to add different types of noise into vowels in the range of 0 dB to 20 dB. Furthermore robustness of new set of feature has been evaluated in the sensor mismatch condition. The classification results show that under noisy background as well as the sensor mismatch condition the proposed technique achieves a better performance over standard cepstral based features. Key-Words: MFCC, GFCC, Formant, HMM, Phoneme Classification.