| Nomaz Mirzaev, Johongir Urinboev, Sayyora Ibragimova, Gulmira Mrzaeva, Azizbek Tillavoldiev The Problem of Forming a Feature Space for Speaker Recognition |
|---|
| Abstract. This work examines the issue of person identification based on voice signals. Speech-based biometric systems are widely used in information security, human-computer interaction, and intelligent information systems. However, noise, channel effects, and high dimensionality of features can negatively af-fect the system's accuracy. This study investigated the development of noise-resistant and computationally efficient recognition algorithms based on a voice dataset consisting of 50 English speakers. Acoustic features such as MFCC, LPC, PLP, spectral centroid, energy, and entropy were extracted from speech signals, and a single feature vector was formed through feature fusion of these features. In this research, person recognition methods were thor-oughly studied, with the main focus on two important approaches: feature fusion and feature space reduction using Principal Component Analysis (PCA) and Independent Component Analysis (ICA). Extensive experiments were conducted on various speech datasets characterized by different noise levels and numbers of speakers. The study yielded good results for a single dataset and classifiers. To reduce redundant and mutually correlated parts of features, dimensionality reduction methods - PCA and ICA - were applied to the feature space. The resulting features were evaluated using K-Nearest Neighbors (KNN) and Linear Discriminant Analysis (LDA) classifiers. These results indicate a significant increase in computational speed and speaker recognition efficiency due to the feature fusion and feature space reduction approaches. |
| Keywords: Speaker identification, speaker verification, feature extraction, feature fusion, feature space reduction, PCA, ICA, KNN, LDA |
 Download PDF |
| DOI: https://doi.org/10.54381/itta2026.3.05 |