MELD (Multimodal EmotionLines Dataset)

Multimodal (audio, video, text) emotion recognition corpus of ~13k utterances from ~1.4k multi-party dialogues sampled from the Friends TV series, labelled with seven emotions and three-way sentiment.

• Task: emotion, speaker
• Languages: en
• Hours: 13
• Domain: tv dialog
• License: GPL-3.0
• Homepage: https://affective-meld.github.io
• Paper: https://aclanthology.org/P19-1050/

Recommendation

Strong choice for benchmarking emotion recognition in conversation, especially multi-party dialog with speaker turns; useful when you need aligned audio+text+visual modalities. TV-acted English only, modest scale (~13 h), class imbalance (Neutral dominant); audio carries music/laugh-track artifacts from broadcast tracks.

Getting the data

Obtain from the dataset homepage.

Distributed via the declare-lab/MELD GitHub repo; raw .mp4 splits are downloadable without gating. GPL-3.0 license applies to the corpus.

Suggested processing

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