From the repository root, run:

conda create -n mdd python==3.10.12
conda activate mdd
pip install torch==2.4.1 torchvision==0.19.1 torchaudio==2.4.1 --index-url https://download.pytorch.org/whl/cu121
pip install -r requirements.txt

The statistics script is inside the data/ folder. Run it from there so it can read train.csv and write the JSON outputs:

cd data
python stat.py

This will generate language-specific statistic files such as:

• data_arabic.json
• data_mandarin.json
• data_hindi.json
• data_korean.json
• data_spanish.json
• data_vietnamese.json

These files contain the computed confusion statistics for each language.

The code assumes all audio files referenced by train.csv, dev.csv, and test.csv exist under EN_MDD/WAV/.

• Copy or move all .wav files into EN_MDD/WAV/
• Make sure the Path column values in data/train.csv, data/dev.csv, and data/test.csv match the file names under EN_MDD/WAV/
• If Path includes subdirectories, preserve the same relative structure under EN_MDD/WAV/

The project currently uses a hardcoded audio base path in these files:

• dataloader.py line 32: ./EN_MDD/WAV/
• train.py line 168: ./EN_MDD/WAV/

If your audio files are stored elsewhere, update those paths accordingly.

From the repository root, run:

python train.py

Optional arguments:

python train.py --num_epoch 100 --batch_size 4 --lr 2e-5