Text-Independent Speaker Identification with Glottal Flow and 1D Convolutional Neural Networks

We propose a method for Text-Independent Speaker Identification that consists in searching for the voiced frames of the speech signal, which are the ones emitted while the vocal cords vibrate, then estimate the glottal flow of each voiced frame using our own iterative inverse filtering technique, we use these glottal flows along with their corresponding speaker labels to train a 1D Convolutional Neural Network (1DCNN). Our specific 1DCNN lacks of dense layers, different from the conventional 1DCNN architecture, it is made of all convolutional layers. In the last layer the number of filters equals the number of speakers, and use global maximum pooling to predict the identity of the speaker that uttered each specific voiced frame. For identifying a speaker we also detect all voiced frames in his/her speech; then estimate the glottal flow of each of these voiced frames; use the trained 1DCNN for identifying the speaker each voiced frame belongs to; and finally, use a voting scheme for deciding the identity of the speaker. We first used our method in the English Language Speech Database for Speaker Recognition (ELSDSR) corpus achieving perfect results (100% accuracy), then we used our technique in the TIMIT database which include audio-files from 630 speakers, using all of them in our tests, achieving an accuracy of 99.52%.