The code in this repository is licensed under Apache-2.0 (see
../LICENSE). The data released alongside it
(DuplexGen/duplexgen-corpus, DuplexGen/duplexgen-spoken on the
HuggingFace Hub) is not covered by that license: each scenario's
DuplexGen-generated dialogues and slot annotations inherit the license of
the upstream source dataset that scenario's text dialogues were seeded
from.
| Scenario | Code | Source dataset | License |
|---|---|---|---|
| Socratic teaching | TEA | SocraticLM | Apache-2.0 |
| Planning | PLN | MultiWOZ | MIT |
| Interview | INT | Anthropic Interviewer | CC-BY-4.0 |
| Negotiation | NEG | CraigslistBargain | MIT |
| Persuasion | PER | DailyPersuasion | Apache-2.0 |
| Social chat | SOC | SODA | CC-BY-4.0 |
Every record in duplexgen-corpus (both the dialogues/<CODE>/ and
annotations/<CODE>/ splits) carries a "license" field matching the table
above, so the applicable license travels with the record.
The Persuasion scenario (PER) is seeded from DailyPersuasion
(PersuGPT), released under Apache-2.0. Please retain attribution to
DailyPersuasion when redistributing the PER split.
For convenience, the Stage-5 TTS script ships two small bundles of
LibriSpeech train-clean-100 audio:
- a tiny 12-speaker sample (one utterance per speaker) under
../tts_render/librispeech_samples/soconvert_spoken.pyruns out of the box without downloading the full corpus; - the fixed assistant voice prompt,
../tts_render/prompt_wavs/assistant_en.wav, from a single held-outtrain-clean-100speaker not included in the 12-speaker pool above (see../tts_render/prompt_wavs/PROVENANCE.md).
Both are part of LibriSpeech (Panayotov et al., 2015), released under CC BY 4.0; please retain attribution to LibriSpeech when redistributing them. For paper-scale rendering, download the full corpus from https://www.openslr.org/12.
Only DuplexGen-generated dialogues (spoken-style conversions and
dialogues with generated turn-taking behavior produced by our pipeline, per
../README.md) and our own human slot-level preference
annotations are released. No raw third-party source text (e.g. the
original MultiWOZ, CraigslistBargain, SocraticLM, DailyPersuasion, SODA, or
Anthropic Interviewer records) is redistributed in this repository or in the
released HuggingFace datasets. The one third-party media exception is the
bundled CC-BY-4.0 LibriSpeech audio described above, bundled purely so the
TTS demo runs without an external download. Users who need the raw
upstream source data, for example to reproduce Stage 1 spoken-style
conversion from scratch, must obtain it directly from the original sources
under their respective licenses.