ParaSpeechCaps: 1M+ Paralinguistic Speech Captions with 59 Style Tags

ParaSpeechCaps provides between 1 million and 10 million speech utterance annotations featuring paralinguistic style captions, released by ajd12342 in 2025. The dataset maps audio characteristics to natural language descriptions using a taxonomy of 59 style tags covering pitch, rhythm, emotion, and environmental noise.

Use Cases

• Controllable text-to-speech synthesis using style tags like pitch and rhythm
• Automated speech captioning to generate descriptions of speaker emotion and vocal quality
• Paralinguistic feature extraction for identifying husky or raspy voice types

Strengths

• Scale of 1M to 10M records
• 59 distinct style tags for granular paralinguistic labeling
• Detailed natural language captions describing complex vocal attributes

Limitations

• Restricted to English language content
• Non-commercial license (CC BY-NC-SA 4.0) restricts industrial use
• Geographic bias toward American environments as noted in sample captions

Provenance

Source

Arxiv 2503.04713

Freshness

Last updated November 2025.

Geography

United States