CNVVE: Dataset and Benchmark for Classifying Non-verbal Voice Expressions

The CNVVE Dataset is consists of 950 audio samples encompassing six distinct classes of voice expressions. These expressions were collected from 42 generous individuals who donated their voice recordings for the study. By making the dataset publicly accessible, we hope to facilitate further research and development of computational methods for non-verbal voice-based interactions.

A more detailed description of the dataset is available in the paper.

Download the dataset

The dataset can be downloaded from Figshare.

Raw dataset: https://darus.uni-stuttgart.de/dataset.xhtml?persistentId=doi:10.18419/darus-3897&version=1.0

Cleaned dataset: https://darus.uni-stuttgart.de/dataset.xhtml?persistentId=doi:10.18419/darus-3898

Setting up the Environment

pip install virtualenv
virtualenv CNVVE
activate CNVVE
pip install -r requirements.txt

Steps to train and make the model

1. Create a new folder call it data and copy the raw audio files in it under data/raw folder (You can also download the cleaned data and skip the next step).

2. Run python clean.py in the terminal

   • This script will trim the empty trailing signals and place the cleaned samples under cleaned folder.

3. Run python normalize.pyin the terminal

   • By default this script normalized the data using zero padding technique. Run python normalize.py --mode=padding to normlize the data using padding technique.

4. Run python augment.py in the terminal

   • By by default it uses the padded data. Run python normalize.py --src_root=data/padded --dst_root=data/padded_augmented to augment the padded data.
   • Note: The created virtualenv ought to be active.

5. Run python createmeta.py --mode=padding to create the metadata.csv

6. Run python train.py to start training the model. Note the training modes in training configuration down below.

Training Configurations

Various modes of training can be used by modifying the config.json.

• The variable train_mode has to be modified for
  • tn: Normal training mode.
  • thp: Hyperparameter search training mode, also modify search_space.json accordingly.
  • tkf: Training with kfold evaluation, modify kfold_num for number of folds accordingly.
  • tp: Training in production mode so the dataset is not split.
  • tea: Training in evaluation for augmentation, for this also do the following below steps.
• For training with augmented dataset the following variables have to be set
  • is_aug: This has to point to true.
  • AUDIO_DIR: This has to point to the new augmented dataset path.
  • ANNOTATIONS_FILE: This has to point to the newly generated metadata csv file.

Citing

If you find this dataset useful please cite:

@inproceedings{hedeshy23_interspeech,
  author={Ramin Hedeshy and Raphael Menges and Steffen Staab},
  title={{CNVVE: Dataset and Benchmark for Classifying Non-verbal Voice}},
  year=2023,
  booktitle={Proc. INTERSPEECH 2023},
  pages={1553--1557},
  doi={10.21437/Interspeech.2023-201}
}