NaturalVoices 870-Hour Podcast Speech Dataset for Voice Conversion

An 870-hour voice conversion dataset contains spontaneous, in-the-wild podcast speech. It was curated by JHU-SmileLab in collaboration with the MSP Lab at CMU LTI for training and evaluating emotion-aware voice conversion systems.

Use Cases

• Train emotion-aware voice conversion models using spontaneous speech from the 870-hour podcast collection.
• Evaluate prosody and emotional expression transfer in voice conversion systems on in-the-wild speech data.
• Develop automatic speech recognition systems on spontaneous, emotionally varied podcast audio.
• Benchmark audio classification models for detecting emotional states or speaker characteristics in podcast speech.

Strengths

• Contains 870 hours of audio data, providing substantial scale for model training.
• Features spontaneous, in-the-wild podcast speech, offering realistic and varied emotional content.
• Curated as part of a research collaboration with JHU-SmileLab and CMU LTI, indicating academic rigor.

Limitations

• Dataset size is described only in hours; the exact number of audio files or rows is unknown.
• The specific audio features, metadata columns, and file formats are not provided in the input.
• Potential geographic or demographic bias may exist as the source is primarily US-based podcasts.

Provenance

Source

JHU-SmileLab, in collaboration with MSP Lab at CMU LTI.

Collection Method

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

Freshness

Last updated on 2025-11-11.

Geography

Primarily United States (based on 'Regionus' tag).