Ultra-low-power voice-activity-detector through context- and resource-cost-aware feature selection in decision trees

Voice-activity-detectors (VADs) are an efficient way to reduce unimportant audio data and are therefore a crucial step towards energy-efficient ubiquitous sensor networks. Current VADs, however, use computationally expensive feature extraction and model building algorithms with too high power requirements to be integrated in low-power sensor nodes. To drastically reduce the VAD power consumption, this paper introduces a decision tree based VAD with (1) a two-phase VAD operation to maximally reduce the power-hungry learning phase, (2) a scalable analog feature extraction block, and (3) context- and dynamic resource-cost-aware feature selection. Evaluation of the VAD was performed with the NOIZEUS database, demonstrating a comparable performance to SoA VADs such as Sohn and Ramírez, while reducing the feature extraction power consumption up to approximately 200 fold.