High Accuracy Silent Speech BCI Using Compact Deep Learning Model for Edge Computing
High Accuracy Silent Speech BCI Using Compact Deep Learning Model for Edge Computing
复制标题
使用紧凑型深度学习模型进行边缘计算的高精度无声语音 BCI
DOI：
10.1109/bci57258.2023.10078589
复制
发表时间：
2023
期刊：
影响因子：
--
通讯作者：
Tomoya Nemoto and Takahiro Morooka
中科院分区：
文献类型：
--
作者：
Nobuaki Kobayashi;Tomoya Nemoto and Takahiro Morooka
无声言语脑机接口（BCI）系统的研究正在进行中，以促进与失去语言能力或患有失语症的护理对象的沟通。该技术通过脑电图分析来假设被护理者在试图说话时试图传达的内容，而不是实际说话。本研究展示了一种当被护理者在想象元音时进行脑电图分析的方法，以实现无声言语BCI护理支持系统。为了构建一个能够减轻用户负担的系统，构建一个能够以较少的电极通道和紧凑的分类器尺寸实现高准确度的大脑活动模型是必不可少的，并且可以在边缘设备中实现。实验通过在被试左颞叶上连接8个电极，测量每个被试想象5种声音和静音（a/i/u/e/0和静音）时的脑电波，并使用四种方法进行分析：支持向量机、决策树、常用的机器学习算法线性区分分析，以及被广泛研究的深度学习模型长短期记忆（LSTM）和EEGNet。采用4名健康男性受试者（23-24岁）对该系统的性能进行评估。结果表明，LSTM和EEGNet对4个被试的平均分类准确率分别为68.8%和80.9%。
Studies on silent speech brain-computer interface (BCI) systems are underway to facilitate communication with care recipients who have lost their language faculty or suffer from aphasia. This technology hypothesizes what a care recipient is attempting to communicate through electroencephalographic analysis when attempting to speak ‘‘without actually speaking This study shows a method of performing electroencephalographic analysis when a person is imagining a vowel sound to realize a silent speech BCI nursing care support system. To construct a system to reduce the burden on users, constructing a model that can achieve a high accuracy of brain activity with less electrode channels and a compact classifier size that can be implemented in edge devices is essential. Experiments were performed by attaching eight electrodes to the left temporal lobes of the subjects, measuring brain waves when each subject was imagining 5 sounds and a mute, i.e., a/i/u/e/0 and mute and performing analysis using four methods: support vector machine, decision tree, linear discriminant analysis, which is a commonly used machine-learning algorithm, and long short-term memory (LSTM) and EEGNet which are extensively investigated deep learning models. The performance of the system was evaluated with four healthy male subjects (ages 23-24). Consequently, the average classification accuracies of 68.8% and 80.9% for four subjects obtained using LSTM and EEGNet.