This article analyzes the basic principles, stages of operation, and modern approaches of speaker diarization systems. Speaker diarization is the process of determining who spoke and when in an audio recording, and is an integral part of automatic speech recognition (ASR) and speaker identification systems. The article extensively covers the main stages of the speaker diarization system—speech activity detection (VAD), feature extraction, embedding creation, clustering, and post-processing. The study analyzes the effectiveness of feature extraction methods such as MFCC, Filterbank Energies, x-vector, ECAPA-TDNN, as well as clustering algorithms such as Agglomerative Hierarchical Clustering (AHC), Spectral Clustering, and DBSCAN. Traditional i-vector, deep learning-based x-vector, and ECAPA-TDNN embedding methods for speaker diarization are compared and their advantages and disadvantages are shown. In addition, the article studies the importance of adaptive filtering, VAD algorithms, and Least Mean Squares (LMS) filtering techniques in speaker diarization systems. The research results show that speaker diarization systems using modern deep learning models can achieve more accurate and reliable results.