An Interpretable Deep Learning Model for Speech Activity Detection Using Electrocorticographic Signals

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Deep learning models for neural speech decoding are often opaque and computationally intensive.
• Existing methods typically require extensive signal preprocessing for effective speech analysis.

Purpose of the Study:

• To develop an explainable deep learning architecture for neural speech decoding and synthesis.
• To create an end-to-end model that automates parameter tuning and provides interpretable results.
• To enable real-time speech detection from raw brain data.

Main Methods:

• A novel deep learning architecture was designed to learn input bandpass filters directly from data.
• The model extracts task-relevant spectral features, enhancing interpretability.
• Intracranial brain data from a speech task was used to implement and test the model.

Main Results:

• The model causally detects speech presence from raw, unprocessed time samples with high accuracy.
• Performance is comparable or superior to existing methods that need significant preprocessing.
• Learned frequency bands align with established neuroscientific findings on speech processing.

Conclusions:

• The proposed model offers an interpretable and computationally efficient approach to neural speech decoding.
• Explainable feature extraction is key to advancing end-to-end speech decoding architectures.
• The model's real-time capability makes it suitable for online brain-computer interface applications.