A Correlation Profile-Based Adaptive Weighing in Mel-DCT Filter Banks for Voice Activity Detection

In this study, we propose an adaptive scheme to dynamically select filters using a novel correlation profile across the modified Mel-DCT (MMD) filter bank which is used to compute the weighted average of the frequency-domain long-term differential entropy (FLDE) and MMD-FLDE features for voice activity detection (VAD) applications. We conduct extensive experiments using the SWITCH-BOARD corpus and noise samples from the NOISEX-92. leveraging this combination of correlation profile-based Mel-DCT filters and long-term speech characteristics. Our results show that the proposed technique outperforms the FLDE and MMD-FLDE methods, in terms of detection accuracy, speech hit rate and noise hit rate, particularly for stationary and heavy noise cases with an improvement of over five percent in the accuracy for machine gun noise in particular. Moreover, we compare the performance of the proposed technique with the robust VAD (rVAD) algorithm, and the results indicate that the proposed method outperforms the rVAD algorithm in terms of speech hit rate.