CNN Architectures and Feature Extraction Methods for EEG Imaginary Speech Recognition

Area of Science:

• Biomedical Engineering
• Computer Science
• Neuroscience

Background:

• Neural dysfunctions significantly impair speech, impacting daily communication.
• Developing accessible assistive technologies is crucial for individuals with communication challenges.
• Imaginary speech recognition offers a potential solution for non-vocal communication.

Purpose of the Study:

• To design and optimize an intelligent imaginary speech recognition system for low-cost applications.
• To identify optimal parameters for feature extraction and neural network architecture.
• To evaluate system performance on a limited-resource platform.

Main Methods:

• Utilized the Kara One database for phoneme and word recordings.
• Employed frequency domain covariance for feature extraction, outperforming time-domain methods.
• Investigated various window lengths (0.25s, 0.5s, 1s) and convolutional neural network (CNN) architectures.
• Tested on eight subjects, aiming for a subject-independent system.

Main Results:

• Frequency domain covariance with a 0.25s window achieved the best performance.
• A CNN with two convolutional layers (64/128 filters) and a dense layer (64 neurons) yielded optimal results.
• Achieved up to 37% accuracy for 11 phonemes and words.
• The system demonstrated a low-cost profile with a 1.8 ms running time on an AMD Ryzen 7 4800HS CPU.

Conclusions:

• Short-term signal analysis is vital for imaginary speech recognition.
• Complex CNN architectures do not guarantee superior performance.
• The developed system is suitable for low-cost, resource-limited applications, offering a viable communication aid.