Decoding Speech Imagery: A Spectro-Spatial Approach to Electroencephalography Band Power Analysis

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Decoding speech imagery from brain signals is crucial for individuals with speech impairments.
• Challenges include limited data and complex brain activity, hindering accurate decoding in prior research.
• Existing methods often overlook spectro-spatial brain activity aspects.

Purpose of the Study:

• To enhance the accuracy of decoding speech imagery from brain signals, especially with limited datasets.
• To investigate the utility of spectro-spatial features for improved decoding.
• To explore potential applications in speech rehabilitation and assistive technologies.

Main Methods:

• Analysis of both spectral (frequency) and spatial (location) aspects of brain activity.
• Utilized time-frequency representation (TFR) features for machine learning model training.
• Trained models on a public Brain-Computer Interface (BCI) dataset (BCI-DB) and a private dataset (PrimAudio-DB).

Main Results:

• Achieved high decoding accuracies: 98.6% on BCI-DB and 81.7% on PrimAudio-DB.
• Identified prominent speech imagery patterns in the frontal region and Gamma frequency band.
• Exceeded previous benchmarks on the BCI-DB dataset.

Conclusions:

• Spectro-spatial analysis significantly enhances non-invasive speech imagery decoding accuracy, even with small datasets.
• The findings offer valuable insights into speech processing in the brain, including language and semantic differences.
• This research paves the way for more effective speech rehabilitation and advanced assistive communication technologies.