Experimental framework for mel-scaled LP based Bangla speech recognition

This paper deals with the recognition process of Bangla speech. The used database consists of two sets of data — one is for training containing 3824 utterances of Bangla digit sequences of 25 male and 25 female speakers and the other one is test dataset containing 1985 utterances of 26 male and 26 female speakers. The test set is subdivided into four groups such as clean1, clean2, clean3 and clean4. Mel-LPC based front-end has been used to design the front-end, since it incorporate auditory-like frequency resolution. The Mel-LPC is a time-domain feature and computationally efficient. The Mel-LPC based cepstral coefficients are obtained directly from the input speech by using generalized autocorrelation function. In this estimation process bilinear transformation is not required, and frequency warping is obtained by using a first-order all-pass filter instead of unit delay. A detail experimental framework both for front-end and HMM based back-end have been presented in this paper. The final recognition experiments show the satisfactory performance of the developed system. The recognition accuracy are found to be 98.11%, 98.05%, 97.94%, and 97.63%, for test sets clean1, clean2, clean3 and clean4, respectively.