NaturalVoices EVC: Emotion-Balanced Spontaneous Podcast Speech

NaturalVoices EVC is an emotion-balanced subset of an 870-hour podcast speech dataset released by JHU-SmileLab and CMU LTI in 2025. It features spontaneous, in-the-wild English audio recordings designed for training and evaluating emotion-aware voice conversion systems.

Use Cases

• Training emotion-aware voice conversion models using emotion labels to modify affect while preserving speaker identity
• Developing audio classification systems to identify emotional states in unscripted English speech
• Fine-tuning text-to-speech (TTS) engines to replicate the prosody of natural podcast conversations

Strengths

• Derived from an 870-hour parent dataset of spontaneous speech
• Emotion-balanced subset ensures representation across affective states
• Sourced from 'in-the-wild' podcast audio rather than laboratory-controlled environments

Limitations

• Restricted to English (US) language, limiting dialectal diversity
• Spontaneous podcast audio may contain background noise or overlapping speech
• Specific row count for this subset is not provided in the repository metadata

Provenance

Source

JHU-SmileLab and MSP Lab at CMU LTI

Collection Method

Curated from spontaneous, in-the-wild podcast speech recordings

Freshness

Last updated November 2025.

Geography

United States