chappiii/whisper-Amharic-finetuning

Fine-tuned OpenAI Whisper (small) for Amharic speech recognition. Trained on Common Voice and FLEURS, achieving 41% WER. # Whisper Fine-tuning for Amharic Speech Recognition Fine-tuning OpenAI Whisper for Amharic speech recognition using production-ready MLOps practices. ## 🎯 **Overview** This project addresses the critical gap in **Amharic speech recognition** by fine-tuning OpenAI's Whisper model (small) on multiple datasets. Amharic, spoken by 25+ million people in Ethiopia, has limited ASR resources, making this work impactful for language accessibility. ### **📦 Published Dataset** As part of this project, the raw BDU-speech dataset used in the original paper was processed and cleaned into a HuggingFace-ready format. The preprocessing involved normalizing audio files, aligning transcriptions, and structuring the data for seamless integration with the 🤗 Transformers training pipeline. The cleaned dataset is publicly available at `chappM/amharic-bdu-asr`, improving accessibility for low-resource Amharic ASR research. ### **Key Insights from Results** While Whisper-small is a powerful multilingual model, our experiments revealed significant challenges with Amharic out-of-the-box: - **Baseline struggles**: The pretrained model often failed to recognize Amharic speech, producing nonsensical outputs (e.g., "2.5." for complex sentences) and achieving a WER of 231.58%. - **Progressive improvement**: Through iterative fine-tuning, we achieved remarkable results—v2 reduced WER to 69.64% using Common Voice, v3 improved to 41% by incorporating FLEURS, and v4 further reduced WER to **29%** by adding the BDU-speech dataset, representing an **87.5% relative improvement** from baseline. - **Production-ready quality**: The v4 model delivers near-perfect, coherent transcriptions that closely match ground truth, making it ready for real-world deployment despite minor optimization needs for inference speed and fast speech handling. This work shows that while multilingual models provide a strong foundation, combining diverse, high-quality datasets is essential for achieving production-ready …