A Comparative Assessment of Text-independent Automatic Speaker Identification Methods Using Limited Data

Automatic Speaker Identification (ASI) is one of the active fields research in signal processing. Various machine learning algorithms have been used for this purpose. With recent developments hardware technologies and data accumulation, Deep Learning (DL) methods become new state-of-the-art approach several classification identification tasks. In paper, we evaluate performance traditional such as Gaussian Mixture Model-Universal Background Model (GMM-UBM) DL-based techniques Factorized Time-Delay Neural Network (FTDNN) Convolutional Networks (CNN) text-independent closed-set automatic speaker on two datasets with different conditions. LibriSpeech experimental datasets, which consists clean audio signals from audiobooks, collected a large number speakers. The other dataset was prepared by us, has rather limited speech low signal-to-noise-ratio real-life conversations customers agents call center. duration query phase an important factor affecting performances ASI methods. work, CNN architecture proposed short segments. design aims at capturing temporal nature optimum convolutional neural network parameters compared to well-known architectures. We show that CNN-based algorithm performs better dataset, whereas amount data, method outperforms all DL approaches. achieved top-1 accuracy model 99.5% 1-second voice instances dataset.