This study presents an artificial intelligence-driven model for detecting semantically similar emergencies in transcribed ECU 911 calls, utilizing advanced Natural Language Processing techniques. The proposed methodology integrates data preprocessing, KeyBERT-based keyword extraction, K-means clustering, and cosine similarity analysis to classify calls with an accuracy of 84.668% (42,334/50,000 comparisons in the second and third quartiles). By filtering extreme quartiles, the model mitigates outlier bias, thereby enhancing ecological validity. The results demonstrate robust pattern identification, with high-similarity call pairs, exemplified by an index of 0.910 for ambulance requests, thereby validating the approach taken. Future work encompasses operational pilots, the integration of deep learning for bias reduction, and the development of real-time processing modules. Our proposal is scalable to other types of emergencies and languages and is positioned as a transformative tool for optimizing emergency response.