DKSCNN: Deep Kronecker Siamese Convolutional Neural Network enabled speaker identification

Abstract

Speaker identification refers to the process of discerning between different voices within audio recordings or streams. Various factors contribute to the complexity of this task, including differences in frameworks, overlapping of sound events, and multiple sound sources present during recording. These aspects significantly complicate the process of speaker identification. To overcome such complexity, a hybrid Deep Kronecker Siamese Convolutional Neural Network (DKSCNN) method is proposed as a solution for performing speaker identification. Initially, the speech signals are collected from the VoxCeleb dataset and it is fed as an input to the pre-processing step and it is performed by the Gaussian distribution-based method. After preprocessing, feature extraction is done to extract features like spectral centroid, pitch chroma, spectral skewness, Power Spectral Density (PSD), Mel-Scale Frequency Cepstral Coefficients (MFCC), logarithmic band power, Hjorth parameters, and tonal power ratio. Based on the extracted features, the speaker identification is done using hybrid DKSCNN, which is the combination of Deep Kronecker Network (DKN) and Siamese Convolutional Neural Network (SCNN). The proposed speaker identification model attained a better accuracy of 90.682%, a True Positive Rate (TPR) rate of 91.362% and a False Positive Rate (FPR) rate of 0.086. The DKSCNN model significantly improves the accuracy and reliability of speaker identification, achieving remarkable performance metrics. This research contributes to the enhancement of speaker identification technology and addresses the problems of real-world audio environments. Also, this approach ensures that speaker identification can be applied across diverse applications.