Adaptive Sampling Framework
The system begins by sampling the audio stream at a current rate. Each audio segment is analyzed by an ML-based aliasing detector that uses STFT features as input. This module is implemented using an XGBoost classifier trained to distinguish aliased from non-aliased signals. If aliasing is detected, the system increases the sampling rate to preserve signal fidelity. If not, the audio is passed to a frequency analysis module, implemented via FFT, to estimate the dominant frequency components in the signal. If the energy is concentrated in low-frequency bands, the sampling rate is decreased to conserve power. If high-frequency energy is present but aliasing is not detected, the system holds the current sampling rate. This dual-stage decision pipeline allows EfficientMic to adjust its behavior at runtime based on both aliasing risk and frequency content.
Adaptive Sampling Timeline
The figure above shows the behavior of EfficientMic over four consecutive recording windows. In the first window, the system detects no aliasing and maintains the current rate. In the second window, high-frequency content triggers aliasing, which is detected in the third. In the fourth window, EfficientMic increases the sampling rate to mitigate aliasing. This closed-loop pipeline enables low-latency, frame-level decisions that balance audio fidelity with resource efficiency. By framing aliasing detection as a lightweight classification task and deferring rate decisions to an FFT-based module, EfficientMic minimizes power usage while maintaining responsiveness. In stable environments, the system reduces the frequency of adjustments to conserve energy, making it well suited for real-world, always-on acoustic sensing in smart infrastructure.