Real-Time Emergency Call Translation
Built for the Saudi Ministry of Hajj and Umrah: a real-time system that translates speech on both sides of an emergency call across Arabic, English, French, Urdu and Indonesian, using fine-tuned open-source STT, LLM and TTS models, plus a call-analysis module that archives every case.
During Hajj, pilgrims from dozens of countries call emergency lines in languages operators cannot understand.
Audio is streamed over WebSocket, transcribed, translated and re-synthesised in both directions with latency low enough to keep a natural conversation. Models run locally through llama.cpp so the system stays available without external API dependencies. Each finished call is summarised and stored for review by the emergency team.
Tools: open-source STT, LLMs and TTS, fine-tuning, WebSocket, llama.cpp.