Table 2.

Detection metrics for three different pre-processing techniques used before training a convolutional neural network: standard spectrograms, mel-scale frequency banks, and mel-frequency cepstrum coefficients. Also included for comparison is the same metrics for the HMM. The accuracies were based on the detectors’ ability to discern between background noise and gunshots up to 800 m away.

	TPR	FPR	FP per Hour	Precision	Recall	F${}_1$ Score
Spectrograms	0.86	0.13	228.0	0.85	0.86	0.85
Freq Banks	0.80	0.04	80.7	0.97	0.80	0.88
MFCC	0.78	0.10	185.8	0.86	0.78	0.82
HMM	0.64	0.08	145.0	0.92	0.64	0.75