Rethinking the Methods and Algorithms for Inner Speech Decoding and Making Them Reproducible.

Area of Science:

• Neuroscience
• Cognitive Science
• Computer Science

Background:

• Inner speech has been studied for decades, but decoding it non-invasively remains challenging.
• Existing brain-computer interface research for inner speech lacks reproducibility and accessible data/code.

Purpose of the Study:

• To investigate machine learning methods for automatic inner speech decoding using Electroencephalography (EEG).
• To address reproducibility issues by making code and data publicly available.
• To evaluate subject-dependent vs. subject-independent approaches and the impact of preprocessing.

Main Methods:

• Utilized Electroencephalography (EEG) data from a publicly available dataset.
• Applied various deep learning models including Convolutional Neural Network (CNN), Gated Recurrent Unit (GRU), and Long Short-Term Memory Networks (LSTM).
• Investigated preprocessing techniques like Independent Component Analysis (ICA), down-sampling, and filtering.

Main Results:

• Achieved 35.20% accuracy in classifying five vowels and 29.21% in classifying six words using a tuned iSpeech-CNN architecture.
• Demonstrated state-of-the-art performance on the chosen public dataset for word classification.
• Evaluated the impact of preprocessing steps and subject-dependent/independent models.

Conclusions:

• The study successfully decoded inner speech using non-invasive EEG and advanced machine learning.
• Publicly sharing code and data significantly enhances the reproducibility of inner speech decoding research.
• This work provides a reproducible framework and valuable insights for future inner speech detection studies.