Design and Implementation of an Efficient Language Independent Speaker Recognition System

Abstract

In this thesis, an algorithm for noise-robust language-independent speaker recognition system is proposed. The first phase for implementation of the proposed speaker recognition system is speaker database collection. To collect database a framework has been designed and presented in this thesis. A small set of fifty speakers has been identified out of the faculty, staff and students of MSIT from different native places as well as having mother tongue other than English and Hindi. Five different sensors are used to compensate for channel mismatch. Samples are recorded in three different speaking styles viz. reading, conversation, and information exchange to obtain a wide range of features. The paradigm for noise robust language independent speaker recognition system is proposed and implemented in MATLAB. The performance of proposed speaker recognition systems is evaluated and compared utilizing MFCC of speech samples recorded in clean environment. Then performance of speaker recognition system using the same arrangement is evaluated with noisy samples. In noisy conditions, the accuracy is degraded. Hence to inculcate the noise robustness in the speaker recognition system, a relatively new feature known as GFCC is proposed to be used and a noise-robust speaker recognition system is implemented in MATLAB again with the same arrangement of neural networks. The result shows that recognition rate is improved with the proposed language independent speaker recognition system using GFCC in the noisy conditions. newlineFurther research can be done for the use of speaker recognition in forensics field for identification of any person and to produce it as evidence in the court of law. There is a scope for new techniques to be discovered and implemented to enhance the performance of the speaker recognition systems for better optimization of the results. For example, the combination of these two features or with some other feature can be used to obtain better performance. newline newline