679190375/chichewa-asr-finetuning

Fine-tuning Whisper for Chichewa (Nyanja) - Automatic Speech Recognition for Low-Resource African Languages # Chichewa ASR Fine-tuning 🎙️ Fine-tuning OpenAI Whisper for **Chichewa (Nyanja)** — a low-resource Bantu language spoken in Malawi. ## 🎯 Project Goal Build an accurate Automatic Speech Recognition (ASR) system for Chichewa using transfer learning and fine-tuning on limited labeled data. ## 🏗️ Approach - **Base model:** OpenAI Whisper (small / medium) - **Technique:** Fine-tuning with CTC + Cross-Entropy loss - **Dataset:** collected Chichewa speech data - **Evaluation:** Word Error Rate (WER) and Character Error Rate (CER) ## 🚀 Getting Started ## 📊 Results | Model | Dataset | WER | CER | |-------|---------|-----|-----| | Whisper-small (baseline) | Common Voice Chichewa | - | - | | Whisper-small (fine-tuned) | Common Voice Chichewa | - | - | *(Results will be updated as experiments progress)* ## 👥 Team - **Researcher:** Ngongbi Laison Chiabi - **Supervisor:** Dr. Dunstan Matekenya - **Tutor:** Tutor Lizette ## 📚 References - OpenAI Whisper Paper - Mozilla Common Voice - Hugging Face Whisper Fine-tuning Guide ## 📄 License MIT License — see LICENSE for details. EOF