EEG Speech Stimuli (Listening) Decoding Research. Uses Brennan 2019 dataset which covers EEG recordings while listening to the first chapter of Alice in Wonderland.
Novel methods proposed:
- Phoneme prediction using EEG features (mostly envelope related) into TransformerEncoder layer and then using MLP to decode Mel Spectrogram.
- 20% phoneme class accuracy on test set (trained on 1916 segments, tested on 223 segments).
Brennan 2019
33 datasets out of 49 were used in the analysis. 8 out of them were excluded due to low performance on the comprehension quiz.
8 of them come from participants with high noise.
Exact number of seconds of all audio files: 723.54 ~= 12 minutes and 3.54 seconds
- S13.mat
- Justification: Noise was acceptable and comprehension score was 8/8.
Many of the datasets were excluded because the pt's performed badly on comprehension tests or the signal contained too much noise. This is the list of usable datasets:
['S01', 'S03', 'S04', 'S05', 'S06', 'S08', 'S10', 'S11', 'S12', 'S13', 'S14', 'S15', 'S16', 'S17', 'S18', 'S19', 'S20', 'S21', 'S22', 'S25', 'S26', 'S34', 'S35', 'S36', 'S37', 'S38', 'S39', 'S40', 'S41', 'S42', 'S44', 'S45', 'S48']
Out of these, S13 is the first dataset where the participant scored 8/8 out of comprehension and where noise didn't render the data unusable.
- audio.zip
- Audio stimuli files
- datasets.mat
- Meta information covering all datasets
- AliceChapterOne-EEG.csv
- Time alignment of text heard by participants
- S__.mat
- EEG dataset for one participant
- /proc/S__.mat
- Preprocessing and EEG alignment info for one participant