The dataset viewer is not available for this split.
Error code: RowsPostProcessingError
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
🎭 VoxSentinel Emotion Dataset
🛡️ Project Overview This dataset is a core component of the VoxSentinel project. It is specifically designed to train and evaluate robust Speech Emotion Recognition (SER) models. By consolidating four of the most reliable and widely-used emotional speech corpora—CREMA-D, MELD, RAVDESS, and TESS—this collection provides a comprehensive, multi-speaker, and cross-gender foundation for modern affective computing and vocal sentiment analysis.
📊 Dataset Structure The entire dataset is bundled into a single high-performance archive to ensure ease of deployment and consistent versioning:
| Filename | Included Sub-datasets | Description |
|---|---|---|
| VoxSentinel_Emotion_Dataset.zip | CREMA, MELD, RAVDESS, TESS | A unified collection of emotional audio clips covering diverse scenarios. |
🔍 Sub-dataset Breakdown:
- CREMA-D: Crowdsourced Emotional Multimodal Actors Dataset (7,442 clips, 91 actors).
- MELD: Multimodal Emotion Lines Dataset (Audio extracted from the 'Friends' TV series, conversational & realistic).
- RAVDESS: Ryerson Audio-Visual Database of Emotional Speech and Song (Professional actors, high-purity audio).
- TESS: Toronto Emotional Speech Set (2,800 clips, focus on older female voices to balance demographics).
🛠️ Data Quality & Refinement To ensure seamless integration with transformer-based architectures like Wav2Vec2 or Hubert, the following refinements have been applied:
- Unified Format: All audio files have been standardized to a consistent sampling rate (e.g., 16kHz) and mono-channel format to prevent feature misalignment.
- Label Mapping: Emotion categories across different datasets (e.g., 'fear', 'happy', 'sad', 'angry', 'neutral') have been mapped to a unified taxonomy.
- Noise Normalization: Volume levels have been normalized across the different sources to ensure the model focuses on prosody rather than recording volume.
🚀 Usage for FinQuest 2026 This dataset is a private asset for the 2026 FinQuest Competition. It serves as the baseline for the emotional intelligence track, challenging participants to detect nuanced human affect in diverse acoustic environments.
How to load
from datasets import load_dataset
# Note: Access is restricted to authorized collaborators.
# Ensure you have the 'VoxSentinel_Emotion_Dataset.zip' in your data path.
dataset = load_dataset("JesseHuang922/VoxSentinel-Emotion-Dataset")
- Downloads last month
- 11