Modelling and Simulation of Speech and Speaker Recognition of A Language

Abstract

In the present paper, soft computing based techniques: neural networks and hybrid approach have being used for the formation of a noise-free artificial word model (AWM) and vowel-diphthong model (VDM). For the formation of a noise-free AWM and VDM, feedforward topology and backpropagation learning of neural networks have been applied. Before formation of a noise-free AWM and VDM, enhancement, segmentation and preprocessing have been carried out on the original speech signal. Blind signal separation and spectral subtraction methods have been used for the enhancement of speech. Next the enhanced speech signal has been segmented using a technique called adaptive vector quantization (AVQ). Finally the preprocessing has been done for the extraction of speech features, using discrete wavelet transform, Fishers linear discriminant analysis (FLDA), discrete cosine transform, principal component analysis (PCA) and particle filtering method. The authors have postulated the mechanism for the formation of a noise-free AWM and VDM through a proposed algorithm called SCB_AWM_VDM (soft computing based artificial word model and vowel-diphthong model). The practical implementation of the algorithm has been tested using Bengali (Indian) language and the complexity of the proposed algorithm has been computed. In the present work, a noise-free AWM has been formed with a total of 23650 Bengali words, taking into consideration – 22 speakers of varying ages uttering each of the 215 Bengali words five times and for the formation of VDM a total 1000 vowels, taking into consideration – 10 speakers of varying ages uttering each vowels and diphthongs.