DuplexAct
Evaluation audio for interaction-action taking in end-to-end full-duplex spoken language models.
This dump is 3,456 items from AMI, CORAAL, Loqui, and the synthesized conditions.
Candor clips are not included (license). The remaining 1,235 items are rebuilt from your own
copy of Candor with scripts/Candor/rebuild_candor.py in the GitHub repo.
Each config is one Level-2 action; the split is test.
| Config | n | Protocol | Audio columns |
|---|---|---|---|
wait |
370 | Direct replay | input |
backchannel |
294 | Direct replay | input |
ignore |
400 | Direct replay | input |
turn_taking |
445 | Direct replay | input |
stop |
42 | Simulator cue | lead_in, cue |
hold_floor |
729 | Simulator cue | lead_in, cue |
respond_and_recover |
776 | Simulator cue | lead_in, cue |
interrupt_user |
400 | Simulator barge-in | user_speech, branch_acknowledged, branch_ignored |
Unused audio columns for a protocol are empty. condition is the Level-3 id.
dataset is AMI / CORAAL / Loqui / synthesized.
metadata is the sample's metadata.json as a string.
Load
from datasets import load_dataset
ds = load_dataset("LincolnD/DuplexAct", "wait", split="test")
# Audio columns are {path, bytes} (self-contained parquet).
Restore the on-disk layout used by inference (from the GitHub repo):
python scripts/unpack_hf_dataset.py --repo_id LincolnD/DuplexAct --out_dir evaluation/data
License
This dump mixes sources. Each row's dataset field (AMI / CORAAL / Loqui / synthesized)
selects the terms that apply. The YAML license is other because AMI is CC BY 4.0,
CORAAL is CC BY-NC-SA 4.0, and Loqui is redistributed with permission — that mix is
not one Creative Commons license.
| Source | n (this dump) | Official terms | Notes |
|---|---|---|---|
| AMI Meeting Corpus | 512 | CC BY 4.0 (AMI licence) | Attribution required. Clips are modified (segmented) relative to the original recordings. |
| CORAAL | 658 | CC BY-NC-SA 4.0 (User Guide) | Non-commercial research/education only. CORAAL-derived clips must stay ShareAlike (CC BY-NC-SA). Cite CORAAL. |
| Synthesized | 2160 | original to this benchmark | For research use of DuplexAct. |
| Loqui | 126 | Columbia Academic Commons doi:10.7916/D82R3PW9. Redistributed with permission of the Loqui authors. | Selected evaluation clips only. Attribute Passonneau & Sachar (2014). Copyright remains with the original authors. |
| Candor | 0 | original Candor license | Not included. Obtain Candor from BetterUp or TalkBank and run scripts/Candor/rebuild_candor.py (1,235 items). |
CORAAL ShareAlike means you cannot relicense CORAAL rows as CC BY. NonCommercial means CORAAL rows (and any mix that includes them) are not for commercial use unless you obtain a separate CORAAL commercial license.
Suggested source citations:
- AMI: groups.inf.ed.ac.uk/ami/corpus
- CORAAL: Kendall, Tyler and Charlie Farrington. 2023. The Corpus of Regional African American Language. Version 2023.06. Eugene, OR: The Online Resources for African American Language Project. http://oraal.uoregon.edu/coraal — https://doi.org/10.7264/1ad5-6t35
- Loqui: Passonneau, Rebecca; Sachar, Evaneet. 2014. Loqui Human-Human Dialogue Corpus (Transcriptions and Annotations). Columbia University Academic Commons. https://doi.org/10.7916/D82R3PW9
Citation
- Downloads last month
- 54