manasseh-zw/shona-f5-tts

# Shona F5-TTS Clean F5-TTS pipeline for Shona training, identity adaptation, checkpoint audition, inference, dataset publishing, and model publishing. ## Architecture ```text shona-f5-tts/ ├── jobs/ │ ├── modal_f5_job.py │ ├── modal_f5_hf_job.py │ ├── modal_f5_identity_job.py │ ├── modal_f5_infer.py │ ├── modal_f5_checkpoint_audition.py │ ├── modal_publish_f5_to_hf.py │ └── modal_upload_dataset.py ├── f5_runtime/ │ ├── configs/ │ ├── scripts/ │ └── sna_f5_tts/ ├── README.md └── pyproject.toml ``` ## Folder Intent - `jobs/`: the Modal entrypoints you actually run - `f5_runtime/`: bundled F5 configs, trainer helpers, inference code, and lightweight project runtime support The goal is to keep the architecture obvious: - operational entrypoints live together - bundled F5 internals live together - no dependency on a separate mutable local checkout ## Main Flows - `jobs/modal_f5_hf_job.py`: phase 1 full adaptation on the broader Shona dataset from Hugging Face - `jobs/modal_f5_identity_job.py`: phase 2 LoRA identity adaptation on the 150-sample single-speaker dataset - `jobs/modal_f5_infer.py`: normal inference against a final checkpoint - `jobs/modal_f5_checkpoint_audition.py`: one-sample-per-checkpoint listening pass - `jobs/modal_publish_f5_to_hf.py`: publish a final model repo to Hugging Face - `jobs/modal_upload_dataset.py`: publish the 150-sample dataset to Hugging Face ## Example Commands ```bash uv run modal run -d jobs/modal_f5_hf_job.py --epochs 30 uv run modal run -d jobs/modal_f5_identity_job.py --epochs 50 uv run modal run jobs/modal_f5_infer.py uv run modal run jobs/modal_f5_checkpoint_audition.py uv run modal run -d jobs/modal_publish_f5_to_hf.py --hf-username manassehzw --repo-name sna-f5-tts ```