de Borman Aurélie, Wittevrongel Benjamin, Van Dyck Bob, Van Rooy Kato, Carrette Evelien, Meurs Alfred, Van Roost Dirk, Van Hulle Marc M
Laboratory for Neuro- and Psychophysiology, KU Leuven, Leuven, Belgium.
Department of Neurology, Ghent University Hospital, Ghent, Belgium.
J Neural Eng. 2025 Jul 31;22(4). doi: 10.1088/1741-2552/adf2de.
Speech brain-computer interfaces (BCIs) aim to restore communication for individuals who have lost the ability to speak by interpreting their brain activity and decoding the intended speech. As an initial component of these decoders, speech detectors have been developed to distinguish between the intent to speak and silence. However, it is important that these detectors account for real-life scenarios in which users may engage language-related brain areas-such as during reading or listening-without any intention to speak.In this study, we analyze the interplay between different speech modes: speaking, listening, imagining speaking, reading and mouthing. We gathered a large dataset of 29 participants implanted with electrocorticography electrodes and developed a speech mode classifier. We also assessed how well classifiers trained on data from a specific participant transfer to other participants, both in the case of a single- and multi-electrode classifier.High accuracy was achieved using linear classifiers, for both single-electrode and multi-electrode configurations. Single-electrode classification reached 88.89% accuracy and multi-electrode classification 96.49% accuracy in distinguishing among three classes (speaking, listening, and silence). The best performing electrodes were located on the superior temporal gyrus and sensorimotor cortex. We found that single-electrode classifiers could be transferred across recording sites. For multi-electrode classifiers, we observed that transfer performance was higher for binary classifiers compared to multiclass classifiers, with the optimal source subject of the binary classifiers depending on the speech modes being classified.Accurately detecting speech from brain signals is essential to prevent spurious outputs from a speech BCI and to advance its use beyond lab settings. To achieve this objective, the transfer between participants is particularly valuable as it can reduce training time, especially in cases where subject training is challenging.
语音脑机接口（BCIs）旨在通过解读大脑活动和解码预期语音，为失去说话能力的个体恢复沟通能力。作为这些解码器的初始组件，语音检测器已被开发出来，用于区分说话意图和沉默状态。然而，重要的是，这些检测器要考虑到现实生活中的场景，即用户可能在没有任何说话意图的情况下激活与语言相关的脑区，比如在阅读或倾听时。在本研究中，我们分析了不同语音模式之间的相互作用：说话、倾听、想象说话、阅读和口型模拟。我们收集了一个由29名植入皮层脑电图电极的参与者组成的大型数据集，并开发了一种语音模式分类器。我们还评估了在单个参与者的数据上训练的分类器，在单电极和多电极分类器的情况下，向其他参与者转移的效果如何。使用线性分类器在单电极和多电极配置中都实现了高精度。在区分三类（说话、倾听和沉默）时，单电极分类的准确率达到了88.89%，多电极分类的准确率达到了96.49%。表现最佳的电极位于颞上回和感觉运动皮层。我们发现单电极分类器可以在不同的记录部位之间转移。对于多电极分类器，我们观察到二分类器的转移性能比多分类器更高，二分类器的最佳源主体取决于所分类的语音模式。从脑信号中准确检测语音对于防止语音脑机接口产生虚假输出，并将其应用扩展到实验室环境之外至关重要。为了实现这一目标，参与者之间的转移特别有价值，因为它可以减少训练时间，尤其是在受试者训练具有挑战性的情况下。
