Baseline scripts for ComParE2021

Information on each sub-challenge is given in the respective branch.

• CCS - Covid Coughing Sub-challenge
• CSS - Covid Speech Sub-challenge
• PRS - Primates Sub-challenge
• ESS - Escalation Sub-challenge

General Installation

Linux

If you have conda installed (either miniconda or anaconda), you can execute ./install to setup the two virtual environments needed for executing the experiments. You can activate the core or end2you environments with source ./activate core or source ./activate end2you respectively.

Installing openSMILE

openSMILE is needed for the extraction of LLDs and audio functionals. For Linux users, we provide a script (scripts/download_opensmile.sh) which downloads the latest binary release and puts it in the location expected by extraction codes. If you are on another platform, download and unpack the correct binary release to opensmile at the root of this repository. The layout should then look like this:

opensmile/
  |-- bin/
  |-- config/
  |-- ...

Data

Make sure, you are on the correct branch regarding your chosen sub-challenge. Otherwise (with one of the virtual environments activated), checkout the desired branch. Move or copy the data from the challenge package into the project's root directory, such that the dist folder lies on the same level as scripts/ and src. The layout should look like this:

dist/
  |-- wav/
  |-- lab/
  |-- features/
  |-- end2you/

Then, with one of the environments activated, run dvc add dist. The data will then be added to a local cache using dvc.

Source code

src contains the python scripts used to run the baseline pipeline. It includes scripts for feature extraction and running the linear SVM classification.

Scripts

We provide bash scripts for extracting auDeep features and running end2you on the data in scripts/. You can run them on their own, or automatically as part of the experimental pipeline.

Reproducing the baseline

You can either reproduce the whole baseline from scratch or run only the SVM classifications on the provided features sets.

The whole thing

If you installed the conda environments defined in .env-ymls/, you can reproduce the whole experimental pipeline from scratch by making use of dvc. Running

dvc repro

will execute the steps defined in dvc.yaml with parameters from params.yaml. dvc will detect changes to the inputs of any experimental stage and subsequent calls to dvc repro will rerun all affected steps. You can also check the current status of the pipeline with dvc status. If you want to run individual steps manually, you can also inspect dvc.yaml which defines the individual steps of the baseline experiments. Note that the pipeline includes feature extraction and thus might run for a considerable amount of time. If everything goes well, a folder results/ will be created, containing predictions and metrics for every feature type plus the late fusion and metrics.json will be populated with an aggregation of these results. You can also execute

dvc metrics show

to print all results to your terminal.

Reproducing only the results from the provided features

If you just want to reproduce the results from the baseline paper with the extracted features provided in the challenge packages, you can make use of src/svm.py. First, adapt the code. Change:

FEATURE_BASE='./features'

to:

FEATURE_BASE='./dist/features'

provided, you extracted the challenge package as described in Data.

Now, run the script and pass the type of features (subfolder in ./dist/features) as a command line argument:

python -m src.svm [featureType]

For example, to run the classification on auDeep features:

python -m src.svm auDeep

This will train and evaluate the SVM on auDeep features and create a folder results/auDeep containing predictions and metrics.