Design and Implementation of Autonomous Text Independent Speaker Recognition
Athanasios Andriopoulos
Abstract
Nowadays, our life is fixated around social media and smart devices which lead to the birth of the big data. As a result, the concept of security is threatened to a point there is a need to evolve our security measures. A widely accepted idea is to use biometric characteristics of the user as an authentication method. As biometric characteristics we define human traits such as voice, appearance, fingerprint or ear shape that differ from one human to an other. The reason biometric characteristics are a viable solution lies in the effort needed in order to copy all those traits. In this diploma thesis, a text-independent speaker recognition system is going to be designed and implemented. Utilizing a short term spectral characteristic extraction method, the Mel- Frequency Cepstral Coefficients (MFCC) which is considered a state-of-the-art method for speech processing. As a classifier an Artificial Neural Network (ANN) is going to be used in order to classify the extracted characteristics. Furthermore, we used a metric utilizing the number of identified inputs for each user in order make a decision. In Conclusion, in this diploma thesis there are several aspects to examine the system such as the parts used, the desgning process, the algorithm for signal processing to extract the voice features, the training algorithm of our neural network and the evaluation process for our system. So, our results will define if a voice characteristics alone can be used as a security system.