Downloads · 30 days
0
mustafa-ozan-duman/wavlm-transformer-mos-english
wavlm-transformer-mos-english is a machine learning model from mustafa-ozan-duman. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
This repository provides a pre-trained model for Mean Opinion Score (MOS) Generation. Given a directory of English audio files, this model predicts the perceived quality on a scale of 1 (Poor) to 5 (Excellent).
Downloads · 30 days
0
Access
Public
Updated Jul 14, 2026
Repo size
1.3 GB
Likes
0
Public
Click a slice to open those files.
.pth1.3 GB · 100%
From the Hugging Face model README
This repository provides a pre-trained model for Mean Opinion Score (MOS) Generation. Given a directory of English audio files, this model predicts the perceived quality on a scale of 1 (Poor) to 5 (Excellent).
The architecture uses a WavLM (Base) feature extractor followed by a custom Transformer-based regression head.
First, clone the repository and install the necessary dependencies to set up your local environment:
# Clone the repository
git clone [https://huggingface.co/mustafa-ozan-duman/wavlm-transformer-mos-english](https://huggingface.co/mustafa-ozan-duman/wavlm-transformer-mos-english)
cd wavlm-transformer-mos-english
# Install requirements
pip install -r requirements.txt
🚀 How to Generate MOS Labels
We have provided a dedicated script, predict_folder.py, to handle batch generation. This script will scan a directory for .wav files, convert them to the required 16kHz format automatically, and save the predictions to a CSV file.
Basic Usage:
Run the following command in your terminal:
Bash
python predict_folder.py --dir path/to/your/audio_folder --out my_predictions.csv
Script Arguments:
--dir: (Required) Path to the directory containing your audio files.
--out: (Optional) Name of the output CSV file (Default: mos_predictions.csv).
📂 Repository Structure
best_model.pth: The trained model weights (1.33 GB).
model.py: The WavLM + Transformer architecture definition.
dataset.py: Audio preprocessing and loading logic.
predict_folder.py: The main script for generating MOS labels.
requirements.txt: Exact library versions used during development.
📝 Technical Details
Sampling Rate: The model operates at 16,000 Hz. Audios with different sampling rates are automatically resampled by the prediction script.
Input Format: Mono PCM Wav.
Output: A CSV file with two columns: wav_filename and predicted_mos.
Author: Mustafa Ozan Duman
Affiliation: Research Assistant, Bursa Uludag University
Contact: [email protected]
🎓 Citation
If you use this model or code in your research, please cite our paper:
Text Citation:
Mustafa Ozan Duman and Ahmet Emir Dirik. "Evaluating SSL and ViViT Architectures for Cross-Corpus Audio MOS Prediction via LODO Validation." arXiv preprint arXiv:2607.10146 (2026).
BibTeX:
Kod snippet'i
@misc{duman2026evaluatingsslvivitarchitectures,
title={Evaluating SSL and ViViT Architectures for Cross-Corpus Audio MOS Prediction via LODO Validation},
author={Mustafa Ozan Duman and Ahmet Emir Dirik},
year={2026},
eprint={2607.10146},
archivePrefix={arXiv},
primaryClass={eess.AS},
url={[https://arxiv.org/abs/2607.10146](https://arxiv.org/abs/2607.10146)},
}