Text-Independent Speaker Identification for Afaan Oromo Using Deep Learning

Speaker identification is the task of identifying “who is speaking” by the help of machines based on the speaker’s speech signal. Speaker identification is a key biometrics feature that used in industries, including computers, voice dialing devices, banking, databases, and extremely secure areas. Identifying the speaker helps to use the identity of the owner for many purposes. It is needed powerful forensic analytical mechanism to defend our communities from criminal conduct. However, no study has been conducted on the Afaan Oromo language to be used for this purpose. This study is going to solve the security problems in many areas. To overcome this problem, this study developed a speaker identification model based on the Afaan Oromo language. In this study, an open set text independent speaker identification for Afaan Oromo using deep learning models was developed using MFCC by Librosa to extract speaker feature vectors with MLP, CNN, and LSTM. Keras model API is used to build the neural network. Totrain and test the model the researcher prepared Afaan Oromoo dataset that has a total of 2000 speech samples by taking a 100-sec utterance of 20 different speakers. The speech was taken carefully from YouTube by taking into account every similar condition of a speaker to reducethe effect of recording device variation. In the end after training the proposed models achieves 99.10% on MLP,98.80% on CNN, and 87.65% on LSTM. .Finally, by comparison of all the proposed models, the MLP model outperforms the other two models with an a ccuracy of 99.1%. Therefore, this study chose MLP model for Text-independent speaker identification for Afaan Oromo language.