Neural voice activity detection with high-gamma ECoG signal correlation structure using a chronically implanted brain-computer interface in an individual with ALS

Chronically implanted brain-computer interfaces (BCIs) for speech decoding hold promise for individuals with severe motor impairment. A key translational challenge to wider adoption is understanding and mitigating long-term neural signal variability. This study evaluated whether correlation-based features of electrocorticographic (ECoG) high-gamma signals (HG-C) provide greater long-term stability and robustness than high-gamma log-power (HGLP) features for neural voice activity detection (NVAD). We analyzed an open-source dataset from Angrick et al. [1] of an individual with amyotrophic lateral sclerosis performing a syllable repetition task. Long short-term memory (LSTM) models were trained on HGLP or HG-C features and evaluated across sessions separated from training by up to six months. Feature importance was estimated with permutation-based marginal effects analysis and robustness was tested by simulated electrode disconnections. HG-C achieved comparable to superior NVAD performance, with reduced temporal degradation (-8% vs. -23% F1 score decline). HG-C models leveraged distributed correlation structure and remained stable under both random and systematic electrode disconnections. These findings support HG-C as a candidate neural signal representation for speech BCIs which require long-term stability and robustness while maintaining ease of use for users and caregivers.