Bengali phonetics identification using wavelet based signal feature

Abstract

With the advancement of the voice signal processing, speech to text recognition has become an important area of research. Though some efforts are found for the English language, for regional languages like Bengali, Hindi, Guajarati etc. it is very rare or not started yet. Thus objectives of this work is to develop a method to identify isolated Bengali letter/alphabet (Swarabarna and Banjanbarna), from uttered sound. In speech processing, identifying a particular uttered letter consists of two major steps, Speech Feature Extraction and Feature Classification. Studies show that Mel Frequency Cepstral Coefficient (MFCC) give better representation of human auditory system, but at the same time with increased noise, performance of MFCC degrades, which may be reduced by Discrete Wavelet Transform (DWT). Thus MFCC combined with DWT is used as a feature termed as Mel Frequency Wavelet Transform Coefficient (MFWTC) for this work. For experiment, a sound database is developed by uttering of 43 Bengali alphabets {11 Swarabarna and 32 Banjanbarna} by ten speakers, 20 times for each letter. Then these signals are pre-processed to remove the silent portion from both end points followed by applying pre-emphasized filter. Next, MFCC features are extracted from preprocessed signals. These features are then fine-tuned by applying DWT to compute MFWTC features. Not only these feature, Zero Crossing Count(ZCC) are also used independently to compare with this method. Finally these features are used to recognize the Bengali Barnas using different classifiers (BayesNet, NaiveBayes, IB1, LWL, Classification Via Clustering, Dagging, Multi Scheme, VFI, Conjunctive Rule, ZeroR, BFTree and Simple Cart) available in Weka tools. The classification accuracy is measured using 10-fold cross validation method, which shows the average 47.61% and 62.19% for Swarabarna and Banjanbarna respectively.