Reproducibility Materials for Compact CRNN Benchmarking and Low-Frequency False-Alarm Mitigation on SAFE
Year: 2026
Type: software
Abstract: This record provides the reproducibility materials accompanying the study “Compact CRNN Benchmarking and Low-Frequency False-Alarm Mitigation for Acoustic Fall-Event Classification on SAFE: A Multi-Seed Reproducibility Study.” The archive contains executable training, inference, analysis, and CPU-benchmark code; ten trained compact CRNN checkpoints; exact train/validation/test split manifests; training histories; 40 run-level metric records; 5,720 sample-level predictions; paired statistical analyses; figures; example inference outputs; a results workbook; technical documentation; and SHA-256 integrity hashes. Five clean-only and five low-frequency-augmented CRNN models were trained with paired seeds 42, 123, 456, 789, and 1010. The 262,002-parameter clean-only CRNN achieved 99.16% ± 0.59% clean accuracy. Under deterministic 20–500 Hz interference at 0 dB, targeted training augmentation increased mean accuracy from 72.03% to 93.99% and reduced the non-fall false-positive rate from 55.45% to 8.12%. All five paired seeds changed in the same direction; the two-sided exact Wilcoxon p value was 0.0625. The results are restricted to within-SAFE event-clip classification and controlled perturbations. They do not establish external-dataset, real-home, clinical, cross-device, or continuous-monitoring performance. The full SAFE dataset is not included. SAFE is available from https://www.kaggle.com/datasets/antonygarciag/fall-audio-detection-dataset and is subject to its original CC BY-NC-SA 4.0 license.
Authors: He, Qiquan, Chen, Zhipeng, Zhan, Shuming
Institution: Fuzhou University of International Studies and Trade
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204689007