PromptDialog-v2
We are still preparing the v2 dataset and expect to upload the data files within one week.
PromptDialog is an audiobook-derived multi-speaker speech dialogue dataset designed for conversational speech research. Most sentences are paired with natural-language prompt annotations, and the dataset is released for academic use under controlled access.
For those interested in the dataset, we provide a preview subset at PromptDialog_sub, containing 1,000 dialogue speech samples.
Changes from v1
We used Qwen-3.8-27B to incorporate semantic context into the preparation and annotation of PromptDialog-v2. The following changes build on PromptDialog-v1.
- Speaker descriptions. Alongside the original description, v2 provides a filtered version focused on how the current speaker delivers the line. This version leaves out scene details and unrelated descriptions, retaining details of the speaker’s tone, emotion, expression, movement, or voice when they directly relate to the utterance.
- Utterance tags. An utterance can now carry multiple fine-grained tags. The current tag annotations are preliminary.
- Dialogue identification. In v1, we used positional rules to group consecutive sentences into dialogues. In v2, Qwen-3.8-27B uses semantic context to determine whether a sequence of sentences forms a single dialogue.
- Narration in dialogue clips. We corrected some audio clips in which narration was mixed with the dialogue.
- Transcript alignment. We combined automatic speech recognition (ASR) with paralinguistic event detection. Using the ASR transcript, the detected events, and the original novel text, Qwen revised the wording, punctuation, and transcription of paralinguistic events to better reflect the recording.
- Acoustic and semantic correction. The correction step in v1 used acoustic cues alone. In v2, it uses both the acoustic evidence and the meaning of the utterance.
We will describe these changes in more detail in future work. If you have questions about a change or an annotation, please open a discussion on the dataset page.
Access and License
PromptDialog is not an open dataset. Access is gated and requires approval by the dataset maintainers.
Access to PromptDialog is governed by CC BY-NC 4.0 together with the additional PromptDialog Data Use Agreement accepted during the gated access request process. Commercial use, redistribution, public mirroring, resale, sharing with third parties, and publication of modified or derivative versions are prohibited without prior written permission.
By requesting access, users agree to use the dataset only for non-commercial academic research or education, comply with the Data Use Agreement, and cite the associated PromptDialog paper.
Important Notice
The dataset may contain derived audio, text, annotations, or metadata from publicly available source materials. Access approval does not transfer ownership of any third-party content and does not grant permission to use the dataset outside the approved research-only scope.
If you are a rights holder and have concerns about any content, please contact the maintainers for review and takedown.
- Downloads last month
- 8