Paper Contents
Abstract
Emergencies that call for immediate response usually have characteristic auditory distress cues, but most surveillancesystems cannot pick up on human screams in real time. This paper proposes a real-time human scream detection system withdeep learning and rigorous parameter optimization to improve emergency response. The system adoptsaudiomentations-based data augmentation with diverse Gaussian noise amplitudes (0.001-0.050), time stretch ratios(0.4-2.0), and pitch shift ranges (1 to 12 semitones) to enhance model robustness and performance.Depth audio featureextraction employs Mel-Frequency Cepstral Coefficients (MFCCs), chroma, and spectral features using the librosa library.The classification pipeline uses Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) models, withaugmented datasets used to train them to distinguish screams from background noise and speech . Real-time detection isperformed through continuous, overlapping 1-second audio buffers via sounddevice for immediate response. Uponconfirmation of a scream through probability thresholding, multi-modal warnings are issued, including audible alerts,desktop notifications, and SMS using the Twilio API. Parameter tests revealed medium noise enhancement (0.005-0.030)achieves best F1-score (+3.2%) and efficiency in processing (2.8ms delay). The system achieves 94-98% detectioneffectiveness and incorporates a Streamlit dashboard for monitoring, file analysis, and logging with easy-to-use interface,allowing for deployment in public safety, healthcare, and home environments for reliable, automated emergency response.Keywords: Scream Detection, Deep Learning, Real-time Audio Processing, Parameter Optimization, Emergency Response,MFCC,Multi-modalAlerts,Audiomentations.
Copyright
Copyright © 2025 VARSHA M, NIKHITA N MAHALE, VARSHA R NAIK, VINAY MURTHY S, DIVYASHREE S . This is an open access article distributed under the Creative Commons Attribution License.
