Intoxicated Speech Detection

Dataset

Alcohol Language Corpus

Workflow

Models

• CNN: alc_cnn.ipynb

• Pretrained Wav2Vec2: alc_wav2vec2.ipynb

Feature Extraction

• Feature sets: eGEMAPS, COMPARE
• Tool: openSMILE

File structure

├── README.md
├── alc_cnn.ipynb
├── alc_wav2vec2.ipynb
├── cnn_architecture.png
├── feature_extraction_compare.py
├── feature_extraction_egemaps.py
├── file_name_split
│   ├── create_balanced_data.py
│   ├── file_name_split.py
├── load_convert_data.py
├── requirements.txt
└── workflow.png

Data preprocessing

• Acoustic features of all wav files are extracted using openSMILE.
• Once openSMILE is installed, you can navigate to the config files to remove/add features you want to extract.
• The config file for eGEMAPS feature set is, e.g., saved under: anaconda3/envs/TL/lib/python3.10/site-packages/opensmile/core/config/egemaps

Notes

• Json files are converted to h5 files (load_convert_data.py) to be loaded faster in CNN.