# SemanticAudio Code Release

This directory contains the source code snapshot used for the SemanticAudio demo page and paper experiments.

## Main Entry Points

- `train_semantic.py`: trains the text-to-semantic Semantic Planner.
- `train_vae_scaled.py`: trains the semantic-to-acoustic Acoustic Synthesizer.
- `infer_semantic.py`: runs end-to-end text-to-audio generation.
- `flowedit_final.py`: runs semantic-space training-free FlowEdit.
- `eval_audiocaps.py`: evaluates AudioCaps generation.
- `eval_ttabench.py`: evaluates TTABench generation.
- `edit_benchmark_final.json`: final 200-pair editing benchmark used for paper evaluation.
- `benchmark_verification.json`: confirms that `edit_benchmark_final.json` aligns with the packaged final evaluation captions.

## Model Components

- `dit/`: Diffusion Transformer backbone and configs.
- `flow_matching/`: Flow Matching wrappers for semantic and acoustic generation.
- `generator/`: acoustic synthesizer and semantic projection modules.
- `data_loaders/`: dataset utilities.
- `scripts/`: training and evaluation launch scripts.

Checkpoints and raw datasets are not included in this static project page repository.