Candra Bella Vista, Mohammad Izamul Fikri Fahmi, Muhammad Dzaka Murran Rusid, Vivi Nur Wijayaningrum
This study presents a deep learning approach for detecting emergency-related speech in the Indonesian language. The main issue addressed is the lack of an automated system capable of accurately recognizing emergency words such as 'tolong' (help), 'jangan' (don't), 'kecelakaan' (accident), 'maling' (thief), and 'kebakaran' (fire) in real-world environments. A dataset consists of 1,100 audio recordings was developed and preprocessed, including silence removal and normalization. Each recording was manually labeled as either emergency or non-emergency based on its content. Acoustic features were extracted using Mel-Frequency Cepstral Coefficients (MFCC), along with delta and delta-delta coefficients to better capture dynamic temporal information. A Gated Reccurent Unit (GRU) based acoustic model was trained and evaluated using both 10-fold cross-validation and a holdout test set. The results suggest that GRU-based models are wellsuited for real-time emergency keyword detection in lowresource languages. The evaluation results show that the best performing GRU model achieved 97% accuracy and 100% recall on emergency-class samples in the external test set. © 2025 IEEE.
State Polytechnic of Malang, Information Technology, Malang, Indonesia