• create from conda env

conda env create -f /path/to/environment.yml

• manually install dependencies

pytorch
torchaudio
fairseq 
scipy
tqdm 

Note: Recommended to download fairseq from github instead of using pip

git clone https://github.com/pytorch/fairseq
cd fairseq
pip install --editable ./

1. put the wavefiles in data/train
2. create data/training_list.txt and data/validation_list.txt

• the format shoule be as follows:

wavname1.wav,quality_score1,intelligibility_score1
wavname2.wav,quality_score2,intelligibility_score2
...

• rescale the intelligibility scores and quality scores into similar scales
  the scales of intelligibility and quality scores in our experiments are 0.0-5.0 and 1.0-5.0, respectively

• Download pretained SSL model, put the model in "./fairseq" dir. Password:inqssssl

python train.py

python test.py --datadir /path/to/your/test/wav/dir --ckpt /path/to/your/checkpoint/inqss

Use the pretrained model:

• Download the SSL model, put the model in "./fairseq" dir. Password:inqssssl
• Download the pretained InQSS, Password:inqssssl
• (Note: the results in the paper is the average scores of four models, which are trained with different training and validation splits. Therefore, the scores in the paper are different from the results using this pretrained model.)

• "Generalization Ability of MOS Prediction Networks" Github