Construction sites are dangerous places where accidents happen regularly, and workers can get trapped or hurt in spots where it's hard for rescue teams to find them quickly. Standard safety monitoring tools like wearable sensors and GPS often don't work well on construction sites because buildings and structures block their signals, especially indoors. This research introduces a novel real-time scream detection and localization system tailored for construction sites, designed to work even on sites with limited equipment and low resources. Our method uses two machine learning models working together: an Enhanced Convolutional Neural Network combined with Wav2Vec2 to detect human screams from all the background construction noise. To track location, we use the Generalized Cross-Correlation with Phase Transform (GCC-PHAT) algorithm to measure time delay estimations, as sounds reach different microphones, even when echoes bounce around construction spaces. We then apply gradient descent methods to calculate exactly where the person in distress is located, despite all the noise interference. This combined approach cuts down on false alarms while speeding up emergency response through better detection and accurate location finding. Initial tests show the system works well at both spotting distress calls and determining where they're coming from, even with typical construction site noise in the background. This research provides a major improvement in keeping construction workers safe and could be used in other dangerous work settings. Training code, evaluation tools, and the complete system implementation are available at https://github.com/Anmol2059/construction_safety .