📄 ArEEG_Chars: Dataset for Envisioned Speech Recognition using EEG for Arabic Characters

Summary

The paper targets a gap in brain-computer interface research: no public EEG data existed for Arabic characters, while English letters, words, and directions already had several benchmarks. The authors built ArEEG_Chars, which they describe as the first such Arabic resource. Thirty native speakers wore a 14-channel Emotiv EPOC X headset. Each viewed every one of the 31 Arabic letters on a slide for 10 seconds, then imagined it with eyes closed for 10 seconds, with a 20-second rest between letters. Slides were shuffled per session, and participants were asked at several points whether their focus had slipped. This gave 930 recordings, which were cut into 250 ms windows to yield 39,857 segments. The dataset is released on Mendeley Data. The paper also surveys earlier imagined-speech datasets and the preprocessing, feature, and classifier pipelines used on them. No classification results are reported; baseline deep learning models are left as future work. The resource matters because it opens envisioned-speech BCI research to Arabic speakers, including people with communication disabilities.

Key points

• Dataset: ArEEG_Chars covers 31 Arabic characters from 30 participants (21 male, 9 female). Ages run roughly from the mid-teens to the early 60s, with a mean around 26. It is published on Mendeley Data (doi 10.17632/f2bgr7tdtt.1).
• Hardware: Emotiv EPOC X, with 14 electrodes placed according to the 10-20 system (AF3, F7, F3, FC5, T7, P7, O1, O2, P8, T8, FC6, F4, F8, AF4) and CMS/DRL references. Signals were sampled at 2048 Hz, downsampled to 128 Hz, and sent over Bluetooth.
• Protocol:
  • 10 seconds viewing each letter, then 10 seconds imagining it with eyes closed.
  • 20-second pause between letters.
  • Letter order randomized.
  • Participants abstained from caffeine, alcohol, and smoking beforehand.
  • Each session took about an hour per person.
• Volume: 930 raw recordings were segmented into 39,857 windows of 250 ms.
• Related work: The surveyed systems include Kumar et al. (Random Forest, 82.20% on digits, letters, and images), an LSTM on wavelet-scattering features for direction commands (about 92.6% F1), and SVM with CSP on two vowels (71%).

Interesting ideas

• I1. Shuffling the letter slides for each session stops participants from anticipating the next stimulus, so order effects cannot masquerade as signal. — For Mova (learn): shuffle item order when querying LLM teacher tiers or running evaluations, so labels and scores do not depend on the sequence in which examples are presented.
• I2. Asking participants at the start, middle, and end whether they lost focus is a cheap self-report check on label quality during collection. — For Mova (learn): have teacher tiers and witness parsers emit an explicit "unsure" flag per item, and keep flagged silver trees out of training or down-weight them.
• I3. A long rest between trials clears the previous mental image, so one trial does not contaminate the next. — For Mova (answer): rebuild the world of the text from scratch for each test question, and add a negative control that checks answers do not change when the preceding item is swapped.
• I4. Cutting each long recording into short fixed windows multiplies the number of training samples. The catch is that windows from one recording are strongly correlated, so splits should be made by recording or participant. — For Mova (learn): when sentence-level silver data is derived from documents, split train and test by source document, so evaluation is not inflated by near-duplicate context.
• I5. Logging per-contributor metadata (age, gender, education) makes stratified error analysis possible later. — For Mova (learn): store provenance for every silver annotation (which witness parsers agreed, genre, source), so failures can be traced to specific data slices.