Jeremy-Gitau/kikuyu_tts

# Kikuyu TTS A custom Text-to-Speech (TTS) training pipeline for the Kikuyu language (kik), built on the Piper TTS framework using a VITS neural architecture. ## Overview This project trains a Kikuyu voice model by: 1. Downloading the Kikuyu audio dataset from HuggingFace (`google/WaxalNLP`, `kik_tts` split) 2. Preprocessing and phonemizing text using espeak-ng 3. Fine-tuning a Piper medium-quality (22050 Hz) VITS model 4. Exporting the trained model to ONNX for inference ## Requirements - Python 3.7+ - espeak-ng (system dependency for Kikuyu phonemization) - PyTorch 1.11+ - A pretrained Piper base checkpoint at `./pretrained/base.ckpt` Install espeak-ng: ```bash # macOS brew install espeak-ng # Ubuntu/Debian sudo apt-get install espeak-ng ``` ## Installation ```bash # Clone or navigate to the project cd kikuyu_tts # Create and activate virtual environment python -m venv venv source venv/bin/activate # Install training dependencies cd piper/src/python pip install -e . # Build Cython alignment extension bash build_monotonic_align.sh # Install runtime (optional, for inference) cd ../python_run pip install -e . cd ../../.. # Install project-specific dependencies pip install datasets soundfile librosa ``` ## Training Workflow All scripts are run from the `kikuyu_tts/` subdirectory: ```bash cd kikuyu_tts ``` ### Step 1 — Download and prepare the dataset ```bash python prepare_data.py ``` Downloads audio from HuggingFace, resamples to 22050 Hz, cleans text, and writes: - `dataset/wavs/*.wav` — audio files - `dataset/metadata.csv` — pipe-delimited `wavs/XXXXX.wav|text` pairs ### Step 2 — Validate the dataset (optional) ```bash python validate.py ``` Prints duration statistics (average, min, max, total hours) for the first 100 utterances. ### Step 3 — Resample audio (if needed) ```bash python resample.py ``` Ensures all WAV files are exactly 22050 Hz, overwriting files in place. ### Step 4 — Preprocess for training ```bash python preprocess.py …