RAVDESS: Ryerson Audio-Visual Database of Emotional Speech and Song
Livingstone & Russo (Ryerson University)
7,356 recordings of 24 actors performing emotions, in audio, video, and song.
Twenty-four professional actors speak or sing two matched statements in a neutral North American accent, expressing calm, happy, sad, angry, fearful, surprised, and disgusted emotions at two intensities. Each clip comes as audio only, audio-video, or video only. It is a standard training and test set for emotion recognition from voice and face.
Access
- Cost
- FreeFree download from Zenodo; no account needed.
- License
- Open, non-commercial onlyCC BY-NC-SA 4.0
- Sign-up
- No sign-up
- How you get it
- Bulk downloadWAV, MP4, ZIP
What's in it
- One row is
- One row per recording file (speech or song clip)
- Covers
- Recorded in a lab; North American English speakers
- Years
- 2018 (one-time release)
- Updated
- One-time release
- Size
- 7,356 files, about 24.8 GB total
Quality
Carefully controlled, validated by naive raters, and cited in hundreds of papers; not updated since release.
- Acted rather than spontaneous emotion
- Only 24 actors and two scripted sentences
- Non-commercial license
Connects to
- IEMOCAP: Larger emotional dyadic speech corpus with more natural interaction
- CREMA-D: Similar acted emotional speech and video set with more actors
- EmoDB: Earlier German acted emotional speech corpus
Questions it could answer
- Which acoustic features best separate fearful from angry speech?
- Do models trained on RAVDESS transfer to spontaneous speech?