Somali-English Transcription
# Somali-English Multilingual ASR Fine-Tuning (Azure ML)
This project enhances transcription accuracy for Somali-English audio by utilizing a fine-tuned `wav2vec2` multilingual model. Built with HuggingFace Transformers and trained on Azure ML Studio with low-resource data.
## Stack
- Model: `facebook/wav2vec2-large-xlsr-53`
- Platform: Azure ML Studio
- Libraries: HuggingFace Transformers, PyTorch, torchaudio
- Augmentation: speed perturbation, noise injection, resampling
## Results
- Word Error Rate reduced by ~30%
- Domain-specific vocabulary coverage improved
- Downstream topic classification F1 increased by +12%
## Training Pipeline
1. Preprocess raw Somali-English audio
2. Apply data augmentation
3. Fine-tune on Azure using GPU-enabled cluster
4. Track WER, CER via validation scripts