kulemantu/sw-tts

Swahili TTS # Swahili TTS Fine-tuning for Cloud GPU This repository contains the necessary code and configuration to fine-tune the `facebook/mms-tts-swh` model using the Mozilla Common Voice Swahili dataset on a cloud GPU environment (e.g., Google Colab, AWS SageMaker, GCP AI Platform, Kaggle Kernel). It is adapted from the ylacombe/finetune-hf-vits repository. ## Prerequisites * Cloud environment with a CUDA-enabled GPU (e.g., NVIDIA T4, V100, A100). * Python 3.8+ * Git * Hugging Face account and authentication token (Settings -> Access Tokens). * (Optional) Weights & Biases account for experiment tracking (wandb.ai). ## Setup 1. **Clone the repository:** ```bash git clone cd swahili-tts-cloud-gpu ``` 2. **Create a Python virtual environment (Recommended):** ```bash python -m venv venv source venv/bin/activate # On Windows use `venv\Scripts\activate` ``` 3. **Install requirements:** ```bash pip install -r requirements.txt # Verify torch is installed with CUDA support # python -c "import torch; print(torch.cuda.is_available())" ``` 4. **Build Monotonic Alignment Search:** This requires C build tools (e.g., `build-essential` on Debian/Ubuntu). ```bash cd monotonic_align python setup.py build_ext --inplace cd .. ``` 5. **Log in to Hugging Face:** ```bash huggingface-cli login # Enter your token with write permissions ``` 6. **(Optional) Log in to Weights & Biases:** ```bash wandb login # Enter your API key ``` ## Fine-tuning Process 1. **Prepare the Base Model:** Convert the original MMS checkpoint to include the discriminator. This saves the required model files into the `./models/base_model` directory. ```bash python convert_original_discriminator_checkpoint.py --language_code sw --pytorch_dump_folder_path ./models/base_model ``` *(Note: The language code here is `sw`, consistent with the dataset config name).* 2. **Accept Dataset Terms:** You *must* have accepted the terms for the `mozilla-foundation/common_voice_17_0` dataset on the Hugging Fa …