A Model for Text Independent Amharic Speake Recognition

Abstract

Speaker recognition is a biometrics for recognizing individuals by their claimed identity based on the characteristics extracted from the voice. Identifying individuals who are claimed by the voice is a big problem in the areas where voice is used as information source. Text Independent Amharic Speaker Recognition (TIASR) model is developed to recognize (Identify) speaker by using Amharic speech utterance which can solve the problem stated. TIASR model is designed to have two components such as: Front end processing and backend processing component. Front end component performs activities such as speech acquiring, pre- processing, feature extraction and feature normalization. Silence removal and pre-emphasis is the tasks performed during pre-processing. Mel Frequency Cepstral Coefficient (MFCC) feature is selected and Cepstral Mean Normalization (CMN) is applied to normalize the feature. Backend component performs developing the speaker model and training the speaker model to the recognizer. Gaussian Mixture Model (GMM) is followed and K-means algorithm is used to develop a GMM for each speaker and optimized by EM algorithm. The model is trained by 20 speakers (15 males and 5 females), using isolated word utterances of Amharic number. And it is tested by 12 speakers (7 genuine and 5 imposter) and utterance randomly generated four digit Amharic number code was taken from each speaker. Failure to Enroll Rate (FER) is computed to measure speaker enrolment performance and the recognizer performs fully without failure on the training data. FNMR and FMR is computed to evaluate identification and it releases the average error rate of 16.85 and releases average accuracy of 83.15% at different threshold value.