EmoNet is a Python toolkit for multi-corpus speech emotion recognition and other audio classification tasks.

(c) 2021 Maurice Gerczuk, Shahin Amiriparian, Björn Schuller: Universität Augsburg

Please direct any questions or requests to Maurice Gerczuk (maurice.gerczuk at uni-a.de) or Shahin Amiriparian (shahin.amiriparian at uni-a.de).

All dependencies can be installed via pip from the requirements.txt:

pip install -r requirements.txt

It is advisable to do this from within a newly created virtual environment.

The basic commandline is accessible from the repository's basedirectory by calling:

python -m emo-net.cli --help

This prints a help message specifying the list of subcommands. For each subcommand, more help is available via:

python -m emo-net.cli [subcommand] --help

The toolkit can be used for arbitrary audio classification tasks. To prepare your dataset, resample all audio content to 16kHz wav files (e.g. with ffmpeg). Afterwards, you need label files in .csv format that specify the categorical target for each sample in the training, development and test partitions, i.e., three files "train.csv", "devel.csv" and "test.csv". The files must include the path to each audio file in the first column - relative to a common basedirectory - and a categorical label in the second column. A header line "file,label" should be included.

The CLI has a nested structure, i.e., it uses two layers of subcommands. The first subcommand specifies the type of neural network architecture that is used. Here, "cnn" gives access to the ResNet architecture which also includes residual adapters, based on the training setting. Two other options, "rnn" and "fusion" are also included but untested and in early stages of development. The rest of this guide will therefore focus on the "cnn" subcommand. After specifying the model type, two distinct subcommands are accessible: "single-task" and "multi-task", which refer to the type of training procedure. For single task, training is performed on one database at a time specified by its basedirectory and the label files for train, validation and developments:

python -m emo-net.cli -v cnn single-task -t [taskName] --data-path /path/to/task/wavs -tr train.csv -v devel.csv -te test.csv

One additional parameter is needed that defines the type of training performed. Here, the choice can be made between tuning a fresh model from scratch (-m scratch), fully finetuning an existing model (-m finetune), training only the classifier head (-m last-layer) and the residual adapter approach (-m adapters). For the last three methods, a pre-trained model has to be loaded by specifying the path to its weights via -im /path/to/weights.h5. While all other parameters have sensible default values, the full list is given below:

The "multi-task" command line slightly differs from the one described above. The most notable difference is in how the data is passed. Instead of passing individual .csv files for each partition, a directory - "--multi-task-setup" - which contains a folder with "train.csv", "val.csv" and "test.csv" files for each database has to be specified. Additionally, "-t" now is used to specify a list of databases (subfolders of the multi task setup) that should be used for training. As multi-domain training is done in a round-robin fashion, there is no predefined notion of a training epoch. Therefore, an additional option ("--steps-per-epoch") is used to define the size of an artificial training epoch. These additional parameters are also given in the table below.