Multidiagnosis-speech

Code for finetuning of feature-based baselines and Transformer models for classification of mental disorders from speech.

# make virtual env
pip install -r requirements.txt

mkdir data
# download and unzip data to 'data/multi_diagnosis'
wget ...
unzip ..
# make sure metadata file (CleanData4.csv) is in 'data/multidiagnosis.
# if only access to CleanData.csv, run preprocessing/merge_participant_metadata.py

# run preprocessing scripts
bash run_preprocessing.py

# train baseline models
python train_baseline_models.py

# train wav2vec models
bash train_wav2vec.sh

#evaluate models
bash evaluate_baselines.sh
python evaluate_wav2vec_models.py

Audio augmentation

• Using torch-audiomentation
• IR taking from MIT Acoustical Reverberation Scene Statistics Survey. Placed in 'augmentation_files/ir'. Specificy path in 'augmentation_config.yml'