MoloudAs/Kinyarwanda-language-model

# Kinyarwanda-language-model Reproducible pipeline for preparing, training, adapting, and aligning a Montreal Forced Aligner (MFA) acoustic model for Kinyarwanda. Kinyarwanda is spoken by 12M+ people, yet high-quality ASR/aligner resources are limited. This repo shows how to go from **raw Kaggle data → cleaned text + audio → pronunciation dictionary → MFA training/adaptation → alignments**, with everything scripted in Jupyter notebooks. ## Data sources - **Kaggle:**Kinyarwanda Automatic Speech Recognition — Track B (1000h) Dataset released by Digital Umuganda, funded by the Gates Foundation. --- ## Environment Setup ### 1. Install Python - Recommended versions: **Python 3.9 – 3.11** - Use **Conda** or **venv** to keep the environment isolated. - Check your version: ```python --version``` ### 2. Install Montreal Forced Aligner (MFA) For this pipeline, **Conda** is used (the recommended method in MFA docs). ```bash conda create -n aligner -c conda-forge montreal-forced-aligner conda activate aligner mfa version ``` Reminder: Always run ```conda activate aligner``` before using MFA. If you prefer Docker or source install, see the MFA documentation. ### 3. Install Jupyter Notebook To run preprocessing and cleaning scripts interactively: ```bash pip install notebook jupyter notebook ``` ### 4. Install Python Libraries These are needed for the preprocessing scripts and notebooks: ```bash pip install pandas numpy tqdm regex textgrid praat-parselmouth epitrans panphon ``` ### 5. Install FFmpeg (for audio conversion) MFA requires audio in .wav format. If your dataset is in .webm (as in Kaggle), install FFmpeg: - macOS (Homebrew): ``` brew install ffmpeg ``` - Ubuntu/Debian: ``` sudo apt-get install ffmpeg ``` --- ## Pipeline (Notebooks) ### Step 1 – Import & Clean Transcriptions **Notebook: `notebooks/01_import_and_clean.ipynb`** - Convert `train.json` → `train.csv` (≈180k rows, 14 columns) - Define Kinyarwanda orthographic units (letters + multi-character …