Speaker localization is critical technology with applications in smart homes, surveillance, and robotics, particularly for enhancing human-machine interactions. Traditional localization methods typically rely on large sensor arrays, which add hardware complexity and cost. To address this limitation, this work presents an approach that combines an anisotropic acoustic metamaterial antenna (AMA) with a deep-learning-based voiceprint recognition (VPR) algorithm to achieve speaker localization using only a single microphone. The anisotropic modulation of speech signals by our proposed subwavelength AMA is the core mechanism enabling the method, constituting the physical foundation for single-microphone speaker localization. Our approach can achieve an impressive localization accuracy of 99.4% with speaker-independent and text-independent features, demonstrating robustness in distinguishing speakers from various directions. Numerical calculations are developed to confirm the orthogonality of frequency modulation across channels, enabling the extraction of directional information from speech. These modulated data are then processed using a VPR model to localize the speaker. Experiments were conducted using the Zhvoice corpus dataset to train the VPR model across eight directions. The model, based on the ResNet18 architecture, is chosen for its balance between performance and computational efficiency. The proposed system offers a low-complexity high-accuracy solution for speaker localization, with potential applications in smart devices and interactive systems. This research contributes to the field of human-machine interactions, offering a cost-effective and valid localization system for intelligent applications.

Close

PrevNext

View figure in article

• Acoustic modeling
• Acoustic wave phenomena
• Sound detection
• Sound waves
• Acoustic metamaterials
• Solid-state detectors
• Deep learning
• Sound wave techniques