Abstract
This paper evaluates the effectiveness of various distance measures for unsupervised speaker change detection (SCD) using Mel-frequency cepstral coefficients (MFCCs), which capture essential characteristics of the speech signal. The study identifies the most effective distance metrics and thresholds for achieving high F1 scores by optimizing each distance measure for its respective threshold. Experiments were conducted using multiple distance metrics, including Euclidean distance, cosine similarity, Kullback-Leibler (KL) divergence, vector quantization (VQ) distortion, Bray-Curtis dissimilarity, and others. Among them, VQ distortion and Bray-Curtis dissimilarity achieved the highest F1 scores of 89.21 and 89.29, respectively, while Euclidean distance showed comparable results with an F1 score of 89.15. Precision, recall, and accuracy were used as key metrics to assess the performance of each distance measure. The study underscores the importance of selecting suitable distance measures and thresholds to balance precision, recall, and accuracy in SCD systems, enhancing their robustness and reliability in practical applications. These results contribute to the development of more effective methods for speaker change detection.