Dual-Microphone Speech Localization and Enhancement Using a Generalized Sidelobe Canceler With Deep Voice Activity Detection

Robust speech communication in adverse acoustic environments requires high speech quality and intelligibility. This paper presents an integrated dual-microphone processing framework that jointly addresses voice activity detection (VAD), speech source localization, and speech enhancement. The proposed system employs a deep learning–based VAD trained using combined zero-crossing rate, log-likelihood ratio, and energy features to reliably distinguish speech from nonspeech segments under noisy conditions. The detected speech activity is incorporated into a diagonal unloading beamformer to improve the robustness of direction of arrival estimation, which subsequently guides a generalized sidelobe canceler for effective interference suppression. A coherence-function–based Wiener postfilter is further applied to refine residual noise estimation. The proposed framework is evaluated in terms of VAD performance, source localization accuracy, and speech enhancement effectiveness. VAD performance is assessed using accuracy, recall, precision, and F1-score, while localization accuracy is evaluated using root-mean-square error. Speech enhancement performance is evaluated using perceptual evaluation of speech quality and short-time objective intelligibility and compared with several baseline methods. Results obtained from both the simulated and real-recording experiments demonstrate that the proposed framework generally outperforms baseline methods in VAD accuracy, localization robustness, and speech enhancement performance.