Comparative Analysis of Audio Features for Unsupervised Speaker Change Detection

This study analyzes how various audio features influence speaker change detection in different unsupervised approaches. Ten different audio features, including MFCC, mel-spectrogram, spectral centroid, and pitch, etc. were chosen for analysis. Two distinct unsupervised methods were selected for comparing the effectiveness of these feature: the Bayesian information criterion with Gaussian mixture model (BIC-GMM), a model-based approach, and the Kullback-Leibler divergence with Gaussian (KL-GMM), a metric-based approach. In order to gain a prior insight into which features have a significant impact on the SCD problem, we calculate statistic for two probabilities: the probability of speaker change given there is a feature change, and vice versa. To experimentally check the prior insights, a set of experiments carried out to analyze the influence of various audio feature for SCD. The results showed that in both methods, MFCC demonstrates a robust performance, and ranks at the top. Zero crossing rate, chroma and spectral contrast gives a higher performance in BIC-GMM. Mel-spectrogram consistently ranks at the bottom, suggesting it has the least impact on performance across both methods.