Towards a Spoken Language Understanding Approach in a Local Language: Application Case of the Fulfulde Language in Hospital Emergency Situations

The language barrier is a real obstacle in healthcare settings. In regions where local languages predominate, healthcare workers, who mainly speak the official languages (French and English in Cameroon), encounter difficulties in accurately recording the symptoms of patients who only speak their mother tongue, such as Fulfulde.In this work, we propose a cascading spoken language understanding (SLU) approach for Fulfulde. The methodology adopted is based on a multi-stage processing chain: automatic speech recognition (ASR), based on the fine-tuning of pre-trained models such as Whisper; the use of noise reduction techniques to improve robustness in real-world environments; and natural language understanding (NLU) to extract intent from transcripts. The speech recognition module was trained and evaluated on a corpus of audio files annotated in Fulfulde. Experimental results show that the system achieves a \textit{Word Error Rate} (WER) of 31% for automatic speech recognition and an precision of 59% for the classification module based on multi-class logistic regression.The dataset used is based on words from the health domain recorded by native speakers from the Far North region of Cameroon. The proposed approach makes it possible to quickly capture the symptoms expressed orally by patients in the local language in order to help healthcare personnel establish an appropriate diagnosis.