JB3Ai/Colab_Sepedi_TTS

# Sepedi Piper TTS Colab Pipeline Clean Colab training pipeline for a Sepedi/Northern Sotho Piper TTS voice model. This repo exists to stop the repeated fragile setup loop and preserve a working path: 1. Prepare Python 3.10 and pip. 2. Install the full audio/ML dependency stack in one consolidated pass. 3. Verify CUDA/GPU before training. 4. Clone or reuse Piper without repeatedly deleting patches. 5. Create a top-level `monotonic_align` fallback package. 6. Preprocess the 663-utterance Sepedi dataset using the Setswana (`tn`) proxy. 7. Run a 50-step smoke test before launching long training. ## Current known blockers solved - `ModuleNotFoundError: No module named 'monotonic_align'` - Missing `config.json` after preprocessing - Missing `msgpack`, `pooch`, `decorator`, `scipy`, `numba`, and related audio dependencies - TensorBoard `add_audio` logger issue - Fragile external install of `rhasspy/monotonic-align` ## Important operating rules Do not use the old broken cycle: ```bash rm -rf piper rm -rf piper_train/vits/monotonic_align python3.10 -m pip install git+github.com ``` That flow failed because the external GitHub install could not complete, leaving Piper without any usable `monotonic_align` module. Use the fallback built into this repo instead. ## Colab file Open: ```text colab/Sepedi_Piper_TTS_Training.ipynb ``` or copy the script version: ```text colab/Sepedi_Piper_TTS_Training.py ``` ## Required Colab runtime Before running the notebook: ```text Runtime → Change runtime type → Hardware accelerator → T4 GPU ``` The pipeline intentionally stops if Python 3.10 cannot see CUDA. Do not continue training until this check passes: ```text CUDA available: True GPU: Tesla T4 ``` ## Dataset expectation The notebook expects: ```text /content/drive/MyDrive/sepedi_tts_dataset.zip ``` which should unzip to: ```text /content/sepedi_tts_dataset/ ``` Expected dataset format: ```text metadata.csv w …