Speaker Identification and Spoken word Recognition in Noisy Environment using Different Techniques

In this work, an attempt is made to design ASR systems through software/computer programs which would perform Speaker Identification, Spoken word recognition and combination of both speaker identification and Spoken word recognition in general noisy environment. Automatic Speech Recognition system is designed for Limited vocabulary of Telugu language words/control commands. The experiments are conducted to find the better combination of feature extraction technique and classifier model that will perform well in general noisy environment (Home/Office environment where noise is around 15-35 dB). A recently proposed features extraction technique Gammatone frequency coefficients which is reported as the best fit to the human auditory system is chosen for the experiments along with the more common feature extraction techniques MFCC and PLP as part of Front end process (i.e. speech features extraction). Two different Artificial Neural Network classifiers Learning Vector Quantization (LVQ) neural networks and Radial Basis Function (RBF) neural networks along with Hidden Markov Models (HMMs) are chosen for the experiments as part of Back end process (i.e. training/modeling the ASRs). The performance of different ASR systems that are designed by utilizing the 9 different combinations (3 feature extraction techniques and 3 classifier models) are analyzed in terms of spoken word recognition and speaker identification accuracy success rate, design time of ASRs, and recognition / identification response time .The testing speech samples are recorded in general noisy conditions i.e.in the existence of air conditioning noise, fan noise, computer key board noise and far away cross talk noise. ASR systems designed and analyzed programmatically in MATLAB 2013(a) Environment