ACOUSTIC FEATURES OPTIMIZATION FOR PUNJABI AUTOMATIC SPEECH RECOGNITION SYSTEM

Abstract

With many advances made in automatic speech recognition technology over past few newlinedecades, there is now an increasing demand of developing Indian ASR. There is a huge newlinegap between performance of machines and a human due to lack of resources, complexity newlinein handling feature vectors, decorrelation of feature information and robustness beside newlineever increasing changes in input speech conditions. Different approaches have been newlineexamined to tackle these factors. The aim of the proposed research work is to cope with newlinethese issues through refinement, combination, and integration of front and back end newlineapproaches with different methodologies. newlineOne of the solution to overcome thesis issues is to explore optimization techniques for newlinetraining (GA+HMM, DE+HMM) of large corpora. The proposed method is applied on newlinebaseline acoustic modeling approaches in training stage. We use the stratergies to newlinedevelop them for most frequently used and language resources levied. Punjabi language newlinefalls in this category but for that we need to first build Punjabi speech dataset. So, in this newlinethesis we first build Punjabi speech corpora of isolated and continuous sentences spoken newlineby adult Punjabi speakers. Its performance is not suggested to be productive on large newlinecorpus with traditional approaches at front and back ends of the system. To reduce newlinefeature vector complexity in training stage, Mel Frequency Cepstral Coefficient (MFCC) newlinefeature vectors are combined with optimization algorithms. It refines the processed newlinefeature vectors before performing classification using baseline hidden Markov model newline(HMM) approach. The experiments are then conducted on large vocabulary of Punjabi newlineisolated words. newlineDespite the improvement in performance, a large gap exists due to the mismatch newlinebetween train and test conditions. We try to reduce this gap by using different newlinecombinations of front end approaches (MFCC, Perceptual Linear Prediction (PLP), newlineRelative Spectral Transform (RASTA) - PLP or their fusion). We test them on each newlinerefined modeling approaches through baseline or