Robust Framework for Speaker Independent Tamil Speech Recognition under Noisy Environments using Modified GFCC Features and Machine Learning Techniques

Abstract

The main objective of the research work is to increase the robustness of an Automatic Speech Recognition ASR for Tamil language by introducing an efficient speech front end newlineprocessing techniques The methodology of the proposed work is carried out in three phases In Phase I a framework has been developed with the aid of existing feature extraction newlineand speech recognition techniques for both noise free and noisy data Best techniques have been selected from both types of data based on Word Recognition Rate WRR and Real Time Factor RTF and only the selected techniques are been used in Phase II for achieving further improvements In Phase II the factors affecting the performance of ASR are analyzed and identified The solutions to the identified problems are carefully developed which can be highly suitable for both noise free and noisy environments Five pass pre processing and three modified GFCC features using multi taper Yule Walker AR power spectrum combinational features using formant frequencies combined frequency warping and feature normalizat ion using LPC and Cepstral Mean Normalization CMN are developed The performance improvements of these techniques are assessed initially for noise free data later the robustness of the same proposed techniques are evaluated for different noisy conditions It is proved from the experiments that the proposed techniques are found to be robust and efficient in terms of improving the recognition accuracy for both noise free and noisy conditions In order to increase the performance of noisy speech recognition various speech signal enhancement techniques are implemented in Phase III and they are evaluated using both subjective and objective speech quality measures Based on the outcome the Recursive Least Squares RLS adaptive algorithm is selected and further improved by introducing a reconstruction methodology using Dual Tree Complex Wavelet DTCW Transform Finally the performance of the noisy speech recognition is evaluated before and after applying the RLSDTCW technique