Figure 2
From: Feasibility of decoding covert speech in ECoG with a Transformer trained on overt speech
Decoding performance. Boxplot depicting the distribution of the token error rate (TER) for five combinations of speech for training and testing. Each dot on the plot represents an individual participant. The green boxes represent the TER distribution of the BLSTM seq2seq model, whereas the orange ones represent the Transformer seq2seq model. The blue ones show the Transformer seq2seq models labeled as “Shuffle,” which were trained with the temporally shuffled ECoG signals. We defined outliers as values that are 1.5 times the interquartile range (IQR) away from the median. The IQR represents the range of values within the middle 50% of the data. The significance \((*: p < 0.05, \ \hbox{n.s.}: p \ge 0.05)\) was calculated with the one-side Wilcoxon signed-rank test Holm–Bonferroni corrected for ten comparisons. For each comparison, the effect size using Cohen’s d is included. The numbers in the figure indicate the average ± standard deviation. (a) In-sample evaluation. (b) Out-of-sample evaluation.
