EEG Imagined Speech Recognition

Imagined speech recognition through EEG signals

Installation

The dependencies are listed in the pyproject.toml file.

Install the package in editable mode (recommended):

```shell

Using uv

uv pip install -e .

Or with pip

pip install -e . ```

Additional packages

For Ubuntu: sudo apt-get install graphviz

For macOS (with Homebrew): brew install graphviz

For Windows: Download and install Graphviz from the Graphviz website.

Usage

Configuration file config.yaml

The configuration file config.yaml contains the paths to the data files and the parameters for the different workflows. Create and populate it with the appropriate values. Refer to config-template.yaml.

```yaml

select: classifier: (str) Select one from { RegularClassifier, ClassifierGridSearch, EvaluateClassifier } dataset: (str) Select one from { KaraOne, FEIS } classifier: featuresselectkbest: k: (int | list[int]) scorefunc: (str) Name of the score function to be used for ranking the features before selection. One from { pearsonr, fclassif } modelbasedir: (path) Preferably use files/Models/ models: (list[str]) list of directory names containing the model.py within them. Eg:- [ model-1, model-2, ... ] nsplits: (int) Number of splits in cross-validation. randomstate: (int) Seed value. testsize: (float) Size of test split. trialsize: (float / null) For testing purposes. Use null to use the entire dataset, else this is the fraction of the dataset that will be used. feis: epochtype: (str) One from { thinking, speaking, stimuli } featuresdir: (path) Preferably use files/Features/FEIS/features-1/ rawdatadir: (path) Preferably use files/Data/FEIS/dataeeg/ subjects: (all / list[int] / list[str]) Specify the subjects to be used. Use 'all' to use all subjects. tasks: list[int]) Available tasks:- [0]; Refer utils/feis.py:FEISDataLoader.gettask(); karaone: epochtype: (str) One from { thinking, speaking, stimuli, clearing } featuresdir: (path) Preferably use files/Features/KaraOne/features-1/ filtereddatadir: (path) Preferably use files/Data/KaraOne/EEGdata-1/ lengthfactor: (float) Determines the window length. overlap: (float) Determines the overlap between consecutive windows. rawdatadir: (path) Preferably use files/Data/KaraOne/EEGraw/ subjects: (all / list[int] / list[str]) Specify the subjects to be used. Use 'all' to use all subjects. tasks: (list[int]) Available tasks:- [0, 1, 2, 3, 4]; Refer utils/karaone.py:KaraOneDataLoader.gettask(); tfrdatasetdir: (path) Preferably use files/TFR/KaraOne/tfr_ds-1/ utils: path: (path) Absolute path to the project directory utils folder ```

Classifier model.py

In {classifier.modelbasedir}, create the model.py with the following template.

```python def model(): # Model definition here # Takes certain parameters like random_state from config.yaml return ...

def param_grid(): # Optional. Only useful in ClassifierGridSearch, ignored otherwise. return ...

def resample(): # Optional. Remove/Comment this entire function to disable sampler. # Takes certain parameters like random_state from config.yaml return ...

def crossvalidation(): # Optional. Remove/Comment this entire function to use default CV of 5 splits from StratifiedKFold. # Takes certain parameters like randomstate, n_splits from config.yaml return ...

def pipeline(): # Optional. Remove/Comment this entire function to disable any pipeline functions to be run. ```

Workflows

Run the different workflows using python3 workflows/*.py from the project directory.

1. download-karaone.py: Download the dataset into the {rawdatadir} folder.

2. features-karaone.py, features-feis.py: Preprocess the EEG data to extract relevant features. Run for different epochtypes: { thinking, acoustic, ... }. Also saves processed data as a .fif to {filtereddata_dir}.

3. ifs-classifier.py: Train a machine learning classifier using the preprocessed EEG data. Uses Information set theory to extract effective information from the feature matrix, to be used as features.

4. flatten-classifier.py: Flattens the feature matrix to a vector, to be used as features. Specify the number of features to be selected in featuresselectk_bestk.

5. flatten-classifier-KBest.py: Run over multiple k's from featuresselectk_bestk.

References

• The KARA ONE Database: Phonological Categories in imagined and articulated speech https://www.cs.toronto.edu/~complingweb/data/karaOne/karaOne.html

• https://github.com/scottwellington/FEIS

Citation

If you use this project in your research, please cite using the following BibTeX entries.

bibtex @software{Yedlapalli_EEG-Imagined-Speech-recognition, author = {Yedlapalli, Ashrith Sagar}, license = {MIT}, title = {{EEG-Imagined-Speech-recognition}}, url = {https://github.com/AshrithSagar/EEG-Imagined-speech-recognition} }

License

This project falls under the MIT License.