johnchuma390/kiswahili-stt

AI-Enabled Kiswahili Speech-to-Text System for Education # Kiswahili Speech-to-Text System for Education A lightweight deep learning system that converts spoken Kiswahili into text, designed to run on low-cost hardware (Android/Raspberry Pi) for use in Kenyan classrooms. ## Project Structure ``` kiswahili-stt/ ├── data/ │ ├── raw/ # Downloaded datasets (not committed to Git) │ ├── processed/ # Cleaned and resampled audio │ └── splits/ # Train/validation/test splits ├── notebooks/ # Jupyter notebooks for exploration and training ├── src/ │ ├── data/ # Data loading and preprocessing scripts │ ├── model/ # Model loading and fine-tuning scripts │ ├── evaluate/ # Evaluation and metrics scripts │ └── app/ # Gradio demo application ├── models/ │ ├── checkpoints/ # Training checkpoints (not committed to Git) │ └── quantised/ # Optimised models for edge deployment ├── results/ # Evaluation results and metrics logs └── requirements.txt ``` ## Setup ```bash python -m venv venv source venv/bin/activate pip install -r requirements.txt ``` ## Training Workflow 1. Download and preprocess data: ```bash python src/data/download_datasets.py python src/data/preprocess.py python src/data/prepare_splits.py ``` 2. Run baseline evaluation: ```bash python src/evaluate/baseline_evaluation.py ``` 3. Fine-tune Whisper on Kiswahili: ```bash python src/model/finetune_whisper.py \ --model-name openai/whisper-small \ --output-dir models/checkpoints/whisper-small-sw-ft \ --learning-rate 1e-5 \ --epochs 10 ``` 4. Evaluate fine-tuned checkpoint on test set: ```bash python src/evaluate/finetuned_evaluation.py \ --checkpoint models/checkpoints/whisper-small-sw-ft \ --output results/finetuned_results.json ``` ## Fine-Tune With More Data To increase training data, build an augmented train split by adding cleaned Common Voice Swahili clips to your current FLEURS train split: ```bash python src/data/prepare_augmented_train.py \ --base-splits-dir data/splits \ --common …