1.
Barsha Deka
– Dept. Of Elec. And Commun. Engg., Gauhati Univ. Inst. Of Scie. And Tech., Gauhati Univ., Assam.
2.
Abhishek Dey And S. R. Nirmala
– Dept. Of Elec. And Commun. Engg., Gauhati Univ. Inst. Of Scie. And Tech., Gauhati Univ., Assam.
Abstract
In this work, we present the development of a connected digit recognition system in Assamese language. Assamese is an under-resourced language of North-East India that is widely spoken in the state of Assam. The text corpus used in this work, consists of a sequence 7 digits spoken in continuous manner. In order to capture the variations in phonetic context, the sequence of digits were arranged in such a way that, each digit occur in all the 7 positions. The speech corpus used in this work was collected from 11 native Assamese speakers out of which 5 were female while 6 were male. Mel Frequency Cepstral Coefficient (MFCC) features have been used as front-end features. We have explored the Subspace Gaussian Mixture Model (SGMM) based acoustic modeling approach in addition to the Gaussian Mixture Model (GMM) within the Hidden Markov Model (HMM) framework. Accuracies of 95.7% and 95.9% are achieved in GMM-HMM and SGMM-HMM systems respectively.
