Text Dependent Speaker Recognition Using Artificial Neural Networks For Tigrigna Language Utterance

Abstract

This paper attempted to implement a text-dependent speaker recognition system based on the artificial neural network for Tigrigna language. Speaker recognition is one of the biometric technologies that have been developed almost more than three decades ago. Speaker recognition is a computing task which discriminates between people based upon their voice characteristics. Speaker recognition has divided into two categories based on its tasks: speaker identification and speaker verification. Speaker identification is the process of determining unknown speaker from a known group of speakers. Meanwhile, speaker verification is the process of accepting or rejecting the identity claim of a speaker. Speaker recognition for different languages is still a big challenge for researchers in terms of identification rate accuracy and securities are among the main issues. So, it?�?s vital to find out the root cause of the accuracy performance gap for the speaker model. With the advanced technologies that have been achieved today, researchers have found that using biometric voice as the security system of the efficient one method. The specific activity of speaker recognition methodologies has been used in this thesis. The method divided into two parts. The first part is training process to build a reference model and the second part is testing for the identification process. For identification process, the sound corpus is collected from different speakers considering the terms gender and age taken at sampling frequency 44.1 kHz and16 bit. The development steps of the speaker recognition prototype were dividing into three modules. At first, speech pre-processing techniques such as endpoint detection applied to remove the silent speech. Then, speaker voice characteristics can be extracted using Mel-frequency cepstrum coefficients form feature vector and stored in the database as a reference model. At last, at the stage of pattern matching the unknown inputted speech and the stored reference model are compared using neural network model for the purpose of training the model and identifying the speaker. Performance of the model is measured by parameters of the confusion matrix. The accuracy of the model for text-dependent speaker recognition MFCC based on MLP NN for Tigrigna language is 96.6% and 93.7% for the word and phrase respectively tested by 10 speakers. ANN speaker model showed good promising results. All the tasks were implemented using MATLAB.