AmbiEar: mmWave Based Voice Recognition in NLoS Scenarios

Abstract

Millimeter wave (mmWave) based sensing is a significant technique that enables innovative smart applications, e.g., voice recognition. The existing works in this area require direct sensing of the human’s near-throat region and consequently have limited applicability in non-line-of-sight (NLoS) scenarios. This paper proposes AmbiEar, the first mmWave based voice recognition approach applicable in NLoS scenarios. AmbiEar is based on the insight that the human’s voice causes correlated vibrations of the surrounding objects, regardless of the human’s position and posture. Therefore, AmbiEar regards the surrounding objects as ears that can perceive sound and realizes indirect sensing of the human’s voice by sensing the vibration of the surrounding objects. By incorporating the designs like common component extraction, signal superimposition, and encoder-decoder network, AmbiEar tackles the challenges induced by low-SNR and distorted signals. We implement AmbiEar on a commercial mmWave radar and evaluate its performance under different settings. The experimental results show that AmbiEar has a word recognition accuracy of 87.21% in NLoS scenarios and reduces the recognition error by 35.1%, compared to the direct sensing approach.