Kratos-AI/korean-voice-emotion-dataset

Korean Voice Emotion Dataset

*This dataset contains high-quality (“A-grade”) data. It has been carefully curated, cleaned, and verified to ensure accuracy, completeness, and consistency, making it suitable for high-stakes or production-grade model training.

Dataset Summary

This dataset comprises high-quality Korean speech recordings designed for training and evaluating Speech Emotion Recognition (SER) models. The dataset contains voice samples expressing four distinct emotions: Angry, Happy, Sad, and Surprised. Each recording is categorized by speaker demographics (age: Young/Old, gender: Male/Female), providing a comprehensive resource for emotion classification research in Korean speech.

Contact

For queries or collaborations related to this dataset, contact: - [email protected] - [email protected]

Supported Tasks

• Task Categories: Speech Emotion Recognition (SER)
• Supported Tasks:
  • Emotion classification from speech
  • Audio signal processing for affective computing
  • Speaker demographic analysis
  • Cross-cultural emotion recognition research
  • Voice synthesis with emotional expression (secondary use)

Languages

• Primary Language: Korean (ko)

Dataset Structure

Data Instances

Each instance consists of an audio file and corresponding metadata:

• Audio: WAV format audio files
• Naming Convention: {Age}_{Gender}{ID}_{Emotion}.wav (e.g., Young_Female1_Happy.wav)
• Total Samples: 66 recordings

Data Fields

The dataset includes:

• audio: Audio file in WAV format
• age: Speaker age group (Young or Old)
• gender: Speaker gender (Male or Female)
• emotion: Expressed emotion (Angry, Happy, Sad, Surprised)

Data Distribution

• Emotions: 4 classes (Angry, Happy, Sad, Surprised)
• Demographics:
  • Young speakers: 44 samples (28 Female, 16 Male)
  • Old speakers: 22 samples (12 Female, 10 Male)
• Gender Split: 40 Female, 26 Male recordings

Dataset Creation

Curation Rationale

This dataset was created to advance Korean speech emotion recognition research by providing labeled emotional speech samples across different demographic groups. The inclusion of age and gender metadata enables research into how emotional expression varies across demographics in Korean speech patterns.

Source Data

• Contributors: Korean native speakers across different age groups and genders
• Recording Guidelines: Speakers were asked to express specific emotions naturally in Korean, ensuring authentic emotional expression while maintaining audio quality standards.

Annotations

• Annotation Process: Each audio file is manually labeled with emotion, age group, and gender information
• Annotators: Native Korean speakers familiar with emotional expression patterns
• Quality Control: Multiple validation steps to ensure emotion labels match audio content

Considerations for Using the Data

Social Impact of Dataset

This dataset aims to advance Korean speech emotion recognition technology, potentially improving human-computer interaction in Korean-speaking contexts. Applications include mental health monitoring, customer service automation, and educational technology with culturally-aware emotional understanding.

Discussion of Biases

• Demographic Imbalance: More young speakers (67%) than old speakers (33%), and more female speakers (61%) than male speakers (39%)
• Emotion Representation: Equal distribution across four emotions, but may not reflect natural emotion frequency in real-world scenarios
• Cultural Context: Emotions are expressed within Korean cultural norms, which may differ from other cultures

Other Known Limitations

• Size: Relatively small dataset (66 samples) may limit model generalization
• Emotion Categories: Limited to four basic emotions; complex or mixed emotions not represented
• Audio Quality: Variations in recording conditions may affect model performance
• Regional Dialects: May not represent all Korean regional speech patterns

Intended Uses

✅ Direct Use

• Training and benchmarking Speech Emotion Recognition models for Korean
• Research in cross-cultural emotion recognition
• Development of Korean-language affective computing applications
• Academic research in computational linguistics and psychology

❌ Out-of-Scope Use

• Real-time production systems without additional validation
• Clinical or diagnostic applications for mental health
• Commercial use without proper attribution
• Surveillance or privacy-invasive applications

Considerations and Limitations

• ❗ The dataset is small (66 samples) and may not be fully representative of all Korean speakers
• 💡 Emotional expression is culturally and individually variable — results may differ across listeners
• 🔄 Future versions could benefit from regional dialect inclusion and larger sample sizes
• ⚖️ Demographic imbalances should be considered when training models

License

CC BY 4.0