DistilMOS

Official implementation of DistilMOS: Layer-wise Self-distillation for Self-Supervised Learning Model-based MOS Prediction.

Overview

DistilMOS is a self-supervised learning (SSL) based Mean Opinion Score (MOS) prediction framework that leverages layer-wise self-distillation for robust speech quality estimation.

Installation

git clone https://github.com/BaleYang/DistilMOS.git

cd DistilMOS

conda create -n distilmos python=3.10 -y
conda activate distilmos

pip install -r requirements.txt

Inference

predict.py supports:

• single wav file inference
• directory inference (recursive, batch mode)
• backbone selection: wavlm(default) or w2v2

Single File

python predict.py \
  --input /path/to/audio.wav \
  --ssl_backbone wavlm

Directory (Batch)

python predict.py \
  --input /path/to/wav_dir \
  --ssl_backbone w2v2 \
  --batch_size 32 \
  --output /path/to/predictions.csv

Citation

If you use DistilMOS in your research or project, please cite:

@article{yang2026distilmos,
  title={DistilMOS: Layer-Wise Self-Distillation For Self-Supervised Learning Model-Based MOS Prediction},
  author={Yang, Jianing and Nakata, Wataru and Saito, Yuki and Saruwatari, Hiroshi},
  journal={arXiv preprint arXiv:2601.13700},
  year={2026}
}