Experimental Analysis of an Interactive MFCC + AHC Speaker Diarization Framework Across Multi-Domain Audio Conditions

Automatic Speaker Diarization (ASD)-the process of determining "who spoke when"-is essential for transcription, conversational analytics, call-center monitoring, courtroom recordings, and multilingual human-computer interaction.Classical systems based on MFCCs, GMMs, and hierarchical clustering are interpretable but struggle in noisy, overlapping, and diverse audio conditions, while modern deep-learning approaches like x-vectors, ECAPA-TDNN, and Wav2Vec 2.0 offer higher accuracy but lack transparency.This study evaluates a visualization-enhanced MFCC-GMM-AHC diarization framework across AMI, VoxCeleb, CALLHOME, Mozilla Common Voice, and a custom English-Hindi dataset.The system integrates adaptive VAD, MFCC + Δ + Δ² features, GMM modeling, AHC clustering, and Viterbi re-segmentation with rich diagnostic tools.Results show strong segmentation quality and speaker separability, with DER improving from 12.8% (MFCC-GMM) to 4.7% (Wav2Vec 2.0).The framework demonstrates robust, interpretable, and multidomain performance.