lggvu/avcocktail-turn-taking

AVCocktail Turn-Taking

Turn-taking event labels (speaker shifts and holds) for pairwise speaker conversations, derived from the AVCocktail (from MCoRec challenge) audio-visual cocktail-party dataset.

Files are per-session, per-speaker-pair JSON label files, kept in the same directory layout as the source dataset:

train/
  central-train_channelmaps.pkl # so that the model knows to map to the correct speaker in a dyadic
  session_00/
    events-spk_3-spk_4.json
  session_01/
    events-spk_3-spk_4.json
  ...
dev/
  central-dev_channelmaps_filtered.pkl 
  session_40/
    annotated/
      events-spk_0-spk_1.json
      events-spk_2-spk_3.json
      events-spk_4-spk_5.json
  session_132/
    annotated/
      ...
  ...

Splits

• train (62 files, 56 sessions): turn-taking events produced by a rule-based detector (no human annotation).
• dev (33 files, 15 sessions): turn-taking events manually annotated from Praat TextGrids by a human annotator — used as ground truth for evaluating the rule-based detector / trained models. Only sessions with human annotation available are included; files live under each session's annotated/ subdirectory.

File format

Event labels (events-*.json)

Each JSON file has a single top-level key, roddy_events, containing:

{
  "roddy_events": {
    "shifts": {"0": [<timestamps in seconds>], "1": [<timestamps in seconds>]},
    "holds":  {"0": [<timestamps in seconds>], "1": [<timestamps in seconds>]}
  }
}

• roddy_events is a historical name. Don't ask me why.
• shifts: timestamps where the conversational floor shifts to THIS speaker. If this shift is in speaker 0 list, the timestamps in that list means the moments the floor changed TO speaker 0, not FROM speaker 0. Convetions from other works might differ, you just need to swap the speaker id.
• holds: timestamps where the same speaker holds the floor after a pause.
• Keys "0" and "1" correspond to the first/second speaker named in the filename, e.g. for events-spk_3-spk_4.json, "0" = spk_3, "1" = spk_4 — this also matches the L/R channel order in the channel-map files below.

Channel maps (central-train_channelmaps.pkl, central-dev_channelmaps_filtered.pkl)

Pickled pandas DataFrames, one row per session/speaker-pair, mapping the stereo audio channels to speaker ids:

Column	Description
id	Session + pair identifier, e.g. central-session_23-spk_2-spk_3 (train) or session_54-spk_0-spk_1 (dev)
L	Speaker id on the left (channel 0) — matches the smaller index in the corresponding events-*.json
R	Speaker id on the right (channel 1) — matches the greater index in the corresponding events-*.json

train has 62 rows (one per training pair); dev has 35 rows (one per dev pair).

Notes

• This dataset contains only the turn-taking event labels and channel-mapping metadata, not the underlying audio/video.

License

MIT Licence. We do not own the dataset. We only distribute the labels. The original dataset is AVCocktail (MCoRec). Please check their licence for data use.

Citations

Accepted to IEEE SLT 2026. Citations TBU.