EEG Imagined Speech Recognition

Imagined speech recognition through EEG signals

Original repo can be found here

Completely set up and expanded the code basis

Installation

Follow these steps to get started.

1. Clone the repository:

   git clone https://github.com/wired87/ttt.git
   cd EEG-Imagined-speech-recognition

2. Install dependencies:

   Next, install the required dependencies using pip:

   pip install -r requirements.txt

Linux 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.

---
_select:
  classifier: (str) Select one from { RegularClassifier, ClassifierGridSearch, EvaluateClassifier }
  dataset: (str) Select one from { KaraOne, FEIS }
classifier:
  features_select_k_best:
    k: (int | list[int])
    score_func: (str) Name of the score function to be used for ranking the features before selection. One from { pearsonr, f_classif }
  model_base_dir: (path) Preferably use files/Models/
  models: (list[str]) list of directory names containing the model.py within them. Eg:- [ model-1, model-2, ... ]
  n_splits: (int) Number of splits in cross-validation.
  random_state: (int) Seed value.
  test_size: (float) Size of test split.
  trial_size: (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:
  epoch_type: (str) One from { thinking, speaking, stimuli }
  features_dir: (path) Preferably use files/Features/FEIS/features-1/
  raw_data_dir: (path) Preferably use files/Data/FEIS/data_eeg/
  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.get_task();
karaone:
  epoch_type: (str) One from { thinking, speaking, stimuli, clearing }
  features_dir: (path) Preferably use files/Features/KaraOne/features-1/
  filtered_data_dir: (path) Preferably use files/Data/KaraOne/EEG_data-1/
  length_factor: (float) Determines the window length.
  overlap: (float) Determines the overlap between consecutive windows.
  raw_data_dir: (path) Preferably use files/Data/KaraOne/EEG_raw/
  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.get_task();
  tfr_dataset_dir: (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.model_base_dir}, create the model.py with the following template.

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 cross_validation():
  # Optional. Remove/Comment this entire function to use default CV of 5 splits from StratifiedKFold.
  # Takes certain parameters like random_state, 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 {raw_data_dir} folder.

2. features-karaone.py, features-feis.py: Preprocess the EEG data to extract relevant features. Run for different epoch_types: { thinking, acoustic, ... }. Also saves processed data as a .fif to {filtered_data_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 features_select_k_best[k] (int).

5. flatten-classifier-KBest.py: Run over multiple k's from features_select_k_best[k] (list[int]).

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

License

This project falls under the MIT License.