Multimodal Emotional AI Agent Voice Dataset (ME-AAV) is A Multimodal Feature Dataset for Emotional AI Agent Voice Congnition. This dataset provides a specialized multimodal collection designed for the evaluation of AI agent voices. It integrates acoustic cues, neural responses (EEG), and subjective affective ratings, aligned with real-world popularity indicators. The dataset is organized into three primary directories, each containing specific feature representations and corresponding labels for affective modeling. 1. Directory Structure and Feature Dimensions (1) Audio Features (Acoustic Representations) This folder contains the acoustic characteristics of 80 distinct AI agent voices. Each voice stimulus is represented by a 77-dimensional handcrafted acoustic vector. These features primarily capture paralinguistic differences in intensity, spectral distribution, temporal variation, and voice quality, including: Energy descriptors (3-D), Spectral-flux (3-D), Zero-crossing-rate (3-D), Spectral-centroid (3-D), Spectral-bandwidth (3-D), MFCC statistics (60-D) and Log-Mel descriptors (2-D). (2) EEG Features (Physiological Responses) This folder contains the neural affective responses elicited by the AI agent voices. A 480-dimensional representation was extracted from 32-channel EEG recordings across five frequency bands (delta, theta, alpha, beta, and gamma). For each channel–band pair, three descriptors (absolute power, relative power, and baseline-relative dB change) are provided, representing the listeners' internal physiological reactions to each agent voice. (3) Fused Multimodal Features This folder contains the integrated feature sets for cross-modal analysis. By concatenating the acoustic and physiological vectors, a 557-dimensional multimodal representation was generated. This fusion is designed to evaluate how well a combination of external acoustic properties and internal neural responses can predict the perception of AI agent voices. 2. Affective Labels and Classification Tasks All data entries are annotated with ground-truth labels derived from subjective ratings (34 participants) to support two types of classification tasks: (1) Binary Classification (2-class: A vs. B) The data is partitioned into two categories based on the mean rating threshold of the 10-point scale, representing a transition from low to high levels: Label A (Low): Represents scores in the lower half of the scale (0.0–5.0), indicating "Low Arousal" or "Low Valence." Label B (High): Represents scores in the upper half of the scale (5.0–10.0), indicating "High Arousal" or "High Valence." (2) Four-Level Classification (4-class: A, B, C, D) To provide a more granular mapping of the AI agent voices, the 10-point rating scale is divided into four equal intervals (quartiles), representing a gradient from the lowest to the highest intensity: Label A (Lowest): Represents the first quartile (0.0–2.5), indicating the lowest level of intensity or activation. Label B (Low-Mid): Represents the second quartile (2.5–5.0), indicating a relatively low level of intensity. Label C (Mid-High): Represents the third quartile (5.0–7.5), indicating a relatively high level of intensity. Label D (Highest): Represents the fourth quartile (7.5–10.0), indicating the highest level of intensity or peak activation.