No abstract

The voice sensor is the core part of voice monitoring devices, and it is commonly drifted in long-term running. For this reason, the voice calibration of monitoring devices is essential. Several calibration methods had been introduced by leveraging expensive referred instruments or manual calibration methods. However, these methods are not only dependent on high-cost instruments, but also is impractical on isolated occasions. To overcome these issues, the feature fusion-based neighbor (FbN) model is proposed to calibrate voice sensors, via real-time low-cost ambient sensors. The FbN consists of a real-time awareness stage, feature selection stage, feature fusion stage, and prediction stage. First, voice data and exogenous low-cost sensor (LCS) data are simultaneously collected. Second, those low-cost sensor data are treated as individual features. The irrelevant features are empirically filtered out. The adopted exogenous features are temperature, humidity and air pressure. Third, the selected features are fused to obtain more representative features. Finally, distances between sensor data and represented features are calculated and sorted. The top-[Formula: see text] average distances are regarded as the predictive results. Experimental comparisons with several novelty methods show the effectiveness of the proposed FbN.