Fig. 3: Method and validation results of silent word recognition.
a Pipeline of our deep learning model architecture comprising mainly 3D convolutional layers. b Procedure of evaluating the proposed silent speech recognition system. 100 datasets, comprising 100 words each, are randomly divided into five folds and cross-validated; 58 and 42 datasets out of 100 datasets are acquired from different subjects: A and B, respectively. c Comparison of the recognition performance of two different classifier models, SVM and our deep learning model, as the number of trained data increases. Each accuracy rate is the average value of five independent validations where FOLD 5 in b is fixed as a test dataset, and n datasets randomly selected from the other four folds are trained in our deep learning model. d Word recognition rates in the number of sensor channels. Each accuracy of n channels out of eight channels is the arithmetic mean of the accuracies from all the n-combinations of the eight channels (8Cn) set. e Confusion matrices of word prediction results from three different classifier models, including correlation (left), SVM (middle), and 3D convolution (right), with the average accuracy rates of 10.26%, 76.30%, and 87.53%, respectively.
