The internet now hosts an unprecedented number of audio datasets collected through dedicated acoustic sensors and Internet of Things (IoT) devices deployed across a variety of environments. However, these sensors are often uncalibrated, and their configurations are unknown. As a result, frequency-response variations can distort key audio features, complicating accurate source detection, characterization, and localization. We present a method to estimate the frequency-dependent amplitude response of microphones with unknown characteristics using recordings of signals with known spectral properties. The approach identifies data segments containing signals with known spectral characteristics—such as speech or vehicle noise—and computes their spectra. Aggregating and averaging multiple observations of the same source type improves accuracy and enables uncertainty estimation. Comparing the observed spectra to reference models allows estimation and correction of the microphone’s frequency response. We demonstrate how estimating and correcting the sensor-specific frequency response of IoT acoustic devices enables more accurate source range and source level estimation. By converting previously uncalibrated acoustic data into physical units, we explore the application of this approach to use cases such as vehicle tracking, pattern-of-life analysis, and environmental noise monitoring.
Schnurr et al. (Wed,) studied this question.