"A Multi-Class Electroencephalography Dataset for Imagined Speech Decoding."

"This dataset contains 32-channel electroencephalography (EEG) recordings acquired for multi-class imagined speech decoding using a structured and controlled experimental protocol. Data were collected from ten male participants at the Biomedical Instrumentation Laboratory, Indian Institute of Technology (IIT) Roorkee. Following standard preparation procedures including informed consent, EEG cap placement, and system calibration participants performed tasks across three linguistic categories: vowels, words, and sentences.The signals were recorded using an Emotive Epoch Flex system at a sampling frequency of 128 Hz. The experimental design involved eight distinct stimuli: the vowels \"a\" and \"i\u201d; the words \"water,\" \"sleep,\" \"fine,\" and \"problem\u201d; and the sentences \"I am fine.\" and \"I have problem.\". Each trial followed a fixed temporal structure consisting of a 3-second stimulus presentation, a 4-second imagined speech interval where participants silently imagined articulating the stimulus and a 2-second relaxation phase.To ensure a robust sample size for machine learning and deep learning models, each stimulus was repeated across 60 trials per participant. The raw signals were preprocessed with a 50 Hz notch filter and a 0.1\u201364 Hz bandpass filter to mitigate noise while preserving physiological intent. The dataset is organized hierarchically and provided in MATLAB (.mat) format, enabling reproducible research in signal processing, machine learning, and brain\u2013computer interface (BCI) studies focused on imagined speech."