Aliasing Classification Performance

Table 1 and Figure 5 compare EfficientMic’s aliasing detection module with different machine learning models, including Decision Tree (DT), Random Forest (RF), a dense neural network (DNN), a CNN with convolution and max pooling layers, and a two layer LSTM. The machine learning algorithm chosen for EfficientMic is XGBoost as it outperforms all other models in terms of F1 score while maintaining low energy and latency. EfficientMic consistently achieves better F1-Score for aliasing detection across varying levels of aliasing in ESC-50, UrbanSound8k (Urban), and MAVD datasets, with a minimal time overhead of 0.09 ms. Additionally, EfficientMic consumes less energy than DNN or CNN models, though it
uses more energy than Decision Tree and Random Forest on both the ESC-50 and UrbanSound8K datasets. This demonstrates the system’s ability to achieve the best accuracy without the highest energy consumption.

 

​We evaluate EfficientMic’s time and storage overhead compared to other machine learning models, including CNN, DNN, Decision Tree, and Random Forest. As shown in Figure 5, EfficientMic achieves the lowest inference latency (0.09–0.10 ms), outperforming Random Forest (0.28–0.33 ms), CNN (0.49–0.59 ms), and DNN (1.15–1.16 ms). Including STFT computation, the total latency remains under 20 ms, negligible relative to the 1-second audio window.

​