Gill-tech/kinyarwanda-asr-track-a

# kinyarwanda-asr-track-a # 🇷🇼 Kinyarwanda ASR – Track A This repository contains our solution for the **Kinyarwanda Automatic Speech Recognition (ASR) Hackathon – Track A**, hosted on Kaggle. We develop an ASR model using mel spectrograms, character-level tokenization, and a GRU-based encoder-decoder trained with CTC loss. ## 📁 Structure - `notebooks/` – Jupyter/Kaggle notebooks - `models/` – GRU-based model and training logic - `utils/` – Tokenizer, Dataset, Config files - `train.py` – Optional script for local training - `data/` – Placeholder (data comes from Kaggle competition) ## 🚀 Current Baseline - Model: GRU encoder with CNN front-end - Features: Mel spectrograms (128 dims) - Loss: CTC - Tokenizer: Character-based with special tokens ## 📊 Evaluation We use the official metrics: