Figure 7
From: Decoding of the speech envelope from EEG using the VLAAI deep neural network
Subject-independent VLAAI, trained on the single-speaker stories dataset, finetuned on the subjects of the holdout dataset. Each point in the boxplot is the reconstruction score for a subject, averaged across stimuli. No significant increase is found between the subject-independent model (no finetuning) and models finetuned with 1 and 2 minutes of data. Starting with 5 minutes of available finetuning data, the median reconstruction score seems to increase logarithmically from 0.19 with the amount of training data to 0.25 Pearson r for 120 minutes. (n.s.: p \(\ge\) 0.05, *: 0.01 \(\le\) p < 0.05, **: 0.001 \(\le\) p < 0.01, ***: p < 0.001).
