Downstream Task Performance:
Classifying Audio at different sample rates with CNN

Proposed EfficientMic achieves similar audio class classification performance on ESC-50 and UrbanSound8K with 70.3% lower storage and 40.4% lower energy requirement. Energy requirements and storage needed are the same for both datasets for fixed sampling rates due to similar models and data size during inference. This time and storage includes the time and storage overhead of EfficientMic. 

 

The figures show how sampling rate impacts classification accuracy with a CNN in figure (a), .wav file recording size in figure (b), and STFT energy cost in figure (c). While 44 kHz sampling yields the highest classification accuracy (figure a), it significantly increases storage (figure b) and STFT energy cost (figure c). Lowering the rate to 2 kHz reduces storage by 95.29% and energy by 80.13%, but introduces aliasing that causes a 6.97%–22.5% drop in F1 score. STFT computation time remains under 10 ms across all rates, negligible relative to the 1-second recording duration. EfficientMic achieves 70.28%–70.96% lower storage and 40.39%-51.14% lower energy consumption than the 44 kHz baseline, with only a modest 0.7%–1.16% drop in F1 score. While 16 kHz sampling shows similar performance, it consumes 19.92%–44.38% more energy and 27.6%–30.64% more storage than EfficientMic. Intermediate rates (e.g., 4 kHz and 8 kHz) show linearly scaling overheads but no meaningful F1 gain over 2 kHz, so we report only 2 kHz and 44 kHz in Figure 4 to clearly illustrate the trade-offs. Additionally, we evaluate the performance of EfficientMic on the LSTM architecture. LSTM achieves a 25% higher F1-score in ESC-50 and a 21% lower F1-score in the UrbanSound8K dataset. However, the memory consumption of an LSTM architecture is higher than a CNN model. Thus, we report the results of the CNN architecture. These results confirm that EfficientMic balances classification performance with resource efficiency. Differences in energy and storage impact across the two datasets further reveal how frequency content distribution shapes aliasing effects and system performance.