SPEAKER INDEPENDENT RECOGNITION OF HINDI VOWELS USING THE CONNECTIONIST APPROACH

Abstract

In the work presented, a scheme for speaker independent speech recognition has been implemented. The domain of recognition is selected as a set of Hindi (Devnagari) vowels. To begin, various concepts related to the speech signal, digital processing of speech signal and acoustic-phonetics are dealt ►pith, various phases of .speech recognition are dis-cussed, viz. Sampling, Filtering, Segmentation, Pitch detection and Formants detection. State-of-the-art recognition schemes employing Rule-base (Fuzzy mathematical concepts) and Connectionist architecture (Neural Networks) are explained in brief In the actual in1plenientatioll.schenle, the•speech•is•preenzphasised•with.a-simple filter.--- The preemphasised speech is then segmented by using a simple method which segments the speech to yield distinct phonemes corresponding to different vowels. The Pitch of the speech is then detected using Cepstrum Analysis and the I'ol'lncinl frequencies are extracted hy perfornl-ing Linear Prediction Analysis. A simple neural network, using Backpropagation algorithm, is trained to recognise the Formant frequencies, these being the parameters which completely characterise a vowel. Male and female voice samples, from a wide group of individuals, are collected for neural network training. F, - F2 plot is drawn fbr the data collected as a statistical measure. The software is written in C and C++ Programming languages. The source code is about 3350 lines long. All Miter/ace using X Windom, System Prgranvning has been provided. The programs are run on the Silicon Graphics Workstation 'INDY' in the University Computer Center: