AfolabiDasola/Yoruba-Malaria-ASR

# 🦟 Yoruba Malaria ASR (Wav2Vec2 + KenLM) This project implements an Automatic Speech Recognition (ASR) system tailored for **Yoruba medical speech**, specifically focusing on malaria diagnosis. ## 🔗 Live Demo **Try the model here:** Hugging Face Space *(Click the link to record your voice and test the transcription!)* ## 🎯 Project Goal Malaria is a significant health challenge in Nigeria. Language barriers often hinder effective diagnosis. This model aims to transcribe Yoruba descriptions of symptoms (e.g., *'ibà'*, *'orí fífọ'*) to assist healthcare workers. ## 🛠️ Tech Stack * **Model:** Fine-tuned `facebook/wav2vec2-large-xlsr-53` * **Language Model:** 3-gram KenLM (trained on Yoruba medical corpus) * **Processor:** PyCTCDecode with KenLM integration * **Deployment:** Hugging Face Spaces (Gradio) ## 📂 Repository Structure * `Training_Notebook.ipynb`: The complete code used to fine-tune the model. * `app.py`: The deployment script running on Hugging Face. * `requirements.txt`: Dependencies required to run the inference. ## 🚀 How to Run Locally ```python from transformers import Wav2Vec2ProcessorWithLM, Wav2Vec2ForCTC import librosa import torch # Load the model model_id = "AfolabiDasola/yoruba-malaria-asr-with-lm" processor = Wav2Vec2ProcessorWithLM.from_pretrained(model_id) model = Wav2Vec2ForCTC.from_pretrained(model_id) # Load audio audio, rate = librosa.load("my_audio.wav", sr=16000) # Transcribe inputs = processor(audio, sampling_rate=16000, return_tensors="pt") logits = model(inputs.input_values).logits transcription = processor.decode(logits.cpu().numpy()[0], beam_width=50).text print(transcription)