MSP-IMPROV

Acted dyadic emotional speech corpus from UT Dallas with 12 actors across six dyad sessions producing 8438 speaking turns (652 target sentences) labelled for happiness, sadness, anger, and neutral.

• Task: emotion, speaker
• Languages: en
• Domain: dyadic improvised
• License: see source terms
• Homepage: https://lab-msp.com/MSP/MSP-Improv.html
• Paper: https://ieeexplore.ieee.org/document/7374697/

Recommendation

Use for emotion recognition where you want controlled lexical content but more naturalistic delivery than fully scripted corpora, and for studying perception/elicitation strategies. Good companion to IEMOCAP for cross-corpus generalization. Small speaker pool (12), acted (not in-the-wild).

Getting the data

Obtain from the dataset homepage.

Academic access requires a signed agreement with Prof. Carlos Busso; commercial license is paid (priced at US$8000 per the official page). Total hours not officially published.

Suggested processing

A recommended VoxKitchen pipeline ships in the repository at examples/pipelines/emotion-recognize.yaml — run it with vkit docker run.