Keyword Spotting Using Human Electrocorticographic Recordings.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• Brain-computer interfaces (BCIs) aim to enable communication and control through neural signals.
• Current BCIs often require extensive training and lack the specificity of voice-activated AI.
• Neural keyword spotting offers a potential pathway for low-latency, high-specificity BCIs.

Purpose of the Study:

• To investigate the feasibility of neural keyword spotting using electrocorticography (ECoG) signals.
• To explore the use of motor representations of speech for BCI applications.
• To compare performance based on electrode placement and spatial density.

Main Methods:

• Invasively recorded ECoG signals were used to create neural matched filters for keyword detection.
• Filters were cross-correlated with neural signals to capture temporal activation dynamics.
• Neural vocal activity detection (VAD) and discriminative classifiers identified keyword utterances.

Main Results:

• Performance was strongly influenced by electrode placement and spatial density.
• Vowel discrimination varied significantly between sensorimotor cortex and superior temporal gyrus.
• High-density ECoG in ventral sensorimotor cortex yielded the best keyword detection and VAD performance.

Conclusions:

• Neural keyword spotting is a viable approach for speech BCIs.
• High-density ECoG arrays show potential for achieving wake-word-like functionality.
• Ventral sensorimotor cortex recordings are promising for future BCI development.