---
tags:
  - speech
  - speech-quality-assessment
  - mos-prediction
  - pytorch
---

# SSL Layer MOS checkpoints

This repository contains the projection-head checkpoints associated with
[SSL_Layer_MOS](https://github.com/Hope-Liang/SSL_Layer_MOS), the official
implementation for *Selection of Layers from Self-supervised Learning Models
for Predicting Mean-Opinion-Score of Speech* (IEEE ASRU 2025).

## Contents

The checkpoints are grouped by evaluation dataset:

- `results_nisqa/`
- `results_tencent/`
- `results_bvcc/`

Within each dataset, files are organized by SSL model, layer, and run. Each
experiment includes PyTorch checkpoint files (`.pt`) and its Gin configuration
(`.gin`); logs are retained for reproducibility.

## Usage

Clone the companion code repository and follow its inference instructions:

```bash
git clone https://github.com/Hope-Liang/SSL_Layer_MOS.git
cd SSL_Layer_MOS
python inference.py --audio_path=<YOUR_AUDIO_PATH> --ssl_model=<SELECTED_SSL_MODEL> --ssl_layer=<SELECTED_SSL_LAYER> --ckpt=<PATH_TO_DOWNLOADED_CHECKPOINT>
```

Download a checkpoint from the directory matching your selected dataset, SSL
model, layer, and run.

## Citation

If you use these checkpoints, please cite the associated IEEE ASRU 2025 paper
and the [companion implementation](https://github.com/Hope-Liang/SSL_Layer_MOS).

## License

The license for these checkpoints follows the terms specified by the authors.
Please also comply with the licenses of any underlying SSL models and datasets.