---
license: mit
tags:
  - text-to-articulation
  - articulatory-kinematics
  - speech-synthesis
  - sparc
language:
  - en
datasets:
  - mythicinfinity/libritts_r
---

# STArK (large) — Interspeech 2026 paper checkpoint

This is the checkpoint used for the results reported in
**STArK: Towards Synthesizing Articulatory Kinematics from Text** (Xavier Yin, Carlos Busso,
Interspeech 2026).

- Code: https://github.com/Lab-MSP/STArK
- Paper (temporary link): https://lab-msp.com/MSP/publications/Yin_2026.pdf
- Project page: https://lab-msp.github.io/STArK/

## Which checkpoint is this?

| File / revision | Training step | Notes |
|---|---|---|
| `checkpoint.ckpt` on `main` (tag `paper-step-32k`) | 32,000 | **Paper checkpoint.** `model=large_model train=train_large`, LibriTTS-R `train-clean-100`. |
| `checkpoint.ckpt` at tag `step-100k` | 100,000 | Longer, separate training run. Published here from 2026-08-23 to 2026-11-05. It does **not** match the paper's reported results. |

`main` served the 100k-step checkpoint until 2026-11-05. If you downloaded this model before
that date, you have the 100k-step model, not the paper model. Re-download to get the paper
checkpoint.

SHA-256 of the paper checkpoint (`checkpoint.ckpt` on `main`):
`1a704e045295e90c182001543b7fe0e2cb1fcb6fc9519a932e2bcb2da2dac5f3`

## Usage

```bash
git clone https://github.com/Lab-MSP/STArK && cd STArK && uv sync
python scripts/stark_cli.py synthesize "Hello, this is a test." \
    --reference-audio my_voice.wav --output out.wav
```

The CLI downloads `checkpoint.ckpt` from this repo's `main` branch by default.

To load the 100k-step run instead:

```python
from huggingface_hub import hf_hub_download
path = hf_hub_download("Lab-MSP/stark-large", "checkpoint.ckpt", revision="step-100k")
```

Then pass `--checkpoint <path>` to `stark_cli.py`.

## Citation

```bibtex
@inproceedings{yin2026stark,
  title     = {{STArK}: Towards Synthesizing Articulatory Kinematics from Text},
  author    = {Yin, Xavier and Busso, Carlos},
  booktitle = {Proc. Interspeech 2026},
  year      = {2026}
}
```

STArK resynthesizes speech with the SPARC vocoder; please also cite SPARC
(Cho et al., "Coding Speech through Vocal Tract Kinematics", IEEE JSTSP 2024).