A x-vector based Speaker Recognition in Persian

In this paper, a text-independent speaker recognition system in Persian is implemented by deep neural networks. The x-vector technique based on Time Delay Neural Network (TDNN) is used to extract the embeddings from speech signals. This method attracts researcher’s attention due to noise robustness and high performance. Data augmentation and noise addition are used to improve system performance. The PLDA classifier is used to recognize the speaker. Previous research in the field of “speaker recognition in Persian” is limited. In this work, the network is trained on the Persian part of the CommonVoice dataset. According to the error analysis, non-speech parts of an utterance decrease the accuracy of speaker recognition. So, the non-speech parts are removed by a Convolutional Recurrent Deep Neural Networks (CRDNN). The accuracy of speaker recognition and verification in CommonVoice is 95.24% and 95.56%, respectively. The Equal Error Rate (EER) evaluation metric of the speaker verification system is 4.72%. The attendance monitoring system was developed as one of the applications of the speaker recognition system. System accuracy for 12 and 15 seconds of collected data(includes 16 women and 12 men) is 98.92% and 100%, respectivly.

A x-vector based Speaker Recognition in Persian Keywords:

deep neural networks , speaker recognition , x-vector , Persian language