This is a Python library with utilities for processing emotional speech datasets and training/testing models. There are also command-line tools for common tasks.

This project requires Python 3.7+. It is advised to run the scripts in a Python virtual environment. One can be created with the command

python -m venv .venv

Then you can use this virtual environment:

. .venv/bin/activate

You can install ERTK from PyPI using

pip install ertk

Alternatively you can clone this repository and install using the latest commit:

pip install -r requirements.txt
pip install .

Or, if you want to develop continuously:

pip install -e .

Optional dependencies can be install via:

pip install -r requirements-dev.txt

Or via PyPI:

pip install ertk[all-preprocessors]

Note that if installing from PyPI, fairseq is not updated on PyPI and so must be installed from GitHub directly:

pip install git+https://github.com/facebookresearch/fairseq.git@ae59bd6d04871f6174351ad46c90992e1dca7ac7

Upon installation, you should be able to use common tools using the CLI applications ertk-cli, ertk-dataset and ertk-util. Use the --help option on each one to see what commands are available.

ertk-cli is currently used to run experiments. The exp2 subcommand runs experiments from a config file:

ertk-cli exp2 /path/to/experiment.yaml override1=val1 override2=val2

ertk-dataset has subcommands for viewing and processing datasets. To view info for a dataset, after running the dataset script:

ertk-dataset info corpus.yaml

To view info for individual annotations,

ertk-dataset annotation speaker.csv

See below for use of ertk-dataset process for feature extraction.

ertk-util has miscellaneous utility functions. The most notable is parallel_jobs, which runs multiple experiments in parallel on CPUs or GPUs. Experiments are loaded into a queue and given to the next available worker on a free CPU thread or GPU. The main thread keeps track of failed jobs and writes them to the failed file.

ertk-util parallel_jobs jobs1.txt jobs2.txt --failed failed.txt --cpus $(nproc)

ERTK has several feature extractors and processors built in. There are feature extractors for OpenSMILE, openXBOW, fairseq, huggingface, speechbrain, Keras applications, Audioset models (VGGish and YAMNet), spectrograms, kmeans clustering, resampling, phonemisation, and voice activity detection (VAD) trimming. To list all installed preprocessors, run ertk-dataset process --list_processors.

To run a processor, use ertk-dataset process --features processor. For example, to extract embeddings from the original Wav2vec model:

ertk-dataset process \
    files.txt \
    output.nc \
    --features fairseq \
    model_type=wav2vec \
    checkpoint=/path/to/wav2vec_large.pt

Experiments can be configured with a YAML config file, which specifies the dataset(s) to load, any modifcations to annotations, the features to load, the model type and configuration.

There are processing scripts for many emotional speech datasets in the datasets directory. See datsets/README.md for more information about the supported datasets and the required processing.

See the examples directory for examples.

Our paper has now been publised at ACM Multimedia 2023. If you use ERTK, please cite our paper:

@inproceedings{keesingEmotionRecognitionToolKit2023,
  title = {Emotion {{Recognition ToolKit}} ({{ERTK}}): {{Standardising Tools For Emotion Recognition Research}}},
  shorttitle = {Emotion {{Recognition ToolKit}} ({{ERTK}})},
  booktitle = {Proceedings of the 31st {{ACM International Conference}} on {{Multimedia}}},
  author = {Keesing, Aaron and Koh, Yun Sing and Yogarajan, Vithya and Witbrock, Michael},
  year = {2023},
  month = oct,
  series = {{{MM}} '23},
  pages = {9693--9696},
  publisher = {{Association for Computing Machinery}},
  address = {{New York, NY, USA}},
  doi = {10.1145/3581783.3613459},
}