README.md

---
license: mit
library_name: transformers
pipeline_tag: audio-to-audio
---

# WavCube

WavCube is a 128-dim, 50Hz continuous representation that unifies speech understanding, reconstruction, and generation within a single space. It is presented in the paper [WavCube: Unifying Speech Representation for Understanding and Generation via Semantic-Acoustic Joint Modeling](https://huggingface.co/papers/2605.06407).

- **Code:** [GitHub Repository](https://github.com/yanghaha0908/WavCube)
- **Paper:** [arXiv:2605.06407](https://arxiv.org/abs/2605.06407)

## Usage

Before using the model, ensure you have installed the requirements as described in the [official repository](https://github.com/yanghaha0908/WavCube).

### Extract Representation from Speech
You can get continuous representations from raw wav using the following command:

```bash
python wav_to_feature.py \
    --audio 19_198_000000_000002.wav \
    --config configs/WavCube-stage2.yaml \
    --ckpt WavCube/checkpoints/vocos_checkpoint_epoch=177_step=195000_val_loss=3.3080.ckpt \
    --output 19_198_000000_000002.pt
```

### Reconstruct Speech from Representation
You can reconstruct waveform from representations using the following command:

```bash
python feature_to_wav.py \
    --feature 19_198_000000_000002.pt \
    --config configs/WavCube-stage2.yaml \
    --ckpt WavCube/checkpoints/vocos_checkpoint_epoch=177_step=195000_val_loss=3.3080.ckpt
```

## Citation

```bibtex
@misc{yang2025wavcube,
      title={WavCube: Unifying Speech Representation for Understanding and Generation via Semantic-Acoustic Joint Modeling},
      author={Haohan Yang and others},
      year={2025},
      eprint={2605.06407},
      archivePrefix={arXiv},
      primaryClass={cs.SD},
      url={https://arxiv.org/abs/2605.06407},
}
```