---
license: cc-by-nc-4.0
library_name: pytorch
tags:
- music
- music-structure-analysis
- segmentation
- source-separation
datasets:
- TuneyCompany/tuney-small
---

# PHRASER (Tuney-Small)

Weights for [PHRASER: Loop-Precise Stem-Aware Music Structure Segmentation](https://github.com/library-x/phraser_public),
trained on the Tuney-Small train split (426 tracks). This is the model behind the Tuney-Small
rows of Tables 1 and 2 in the paper.

- File: `phraser_tuney_small.ckpt` (PyTorch dict with `state_dict`, `config`, `epoch`)
- Task network: 31.07M parameters, `PhraserModel(dim_embed=128, num_heads=16)`, on frozen
  MuQ-large features of four SCNet stems. The `dim_embed` stored in `config` is a stale
  default; the network is built with 128.
- Selected by validation loss (epoch 10 of 80).
- SHA-256: see `SHA256SUMS`.

## Usage

    from phraser.modules.phraser import PhraserModel
    import torch
    model = PhraserModel(dim_embed=128, num_heads=16)
    sd = torch.load("phraser_tuney_small.ckpt", map_location="cpu", weights_only=False)["state_dict"]
    model.load_state_dict({k.replace("module.", ""): v for k, v in sd.items()}, strict=False)

The end-to-end evaluation (separation, MuQ, decoding with onset snapping, scoring) is
`python -m phraser.tuney_small.eval_ts phraser phraser_tuney_small.ckpt`. See the repository README.

## Limitations

The model was trained on 426 instrumental, grid-built tracks. It has no vocal supervision, and the
vocal output channel is untrained and must not be used. Accuracy on live or tempo-drifting music is lower.

## License

[CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/). Non-commercial use only. The model
depends on MuQ weights, which are also licensed for non-commercial use only.