RyeAI/coral-v3-conversation-pnc-da

CoRal v3 Conversation PnC DA

RyeAI/coral-v3-conversation-pnc-da is a text-only Danish punctuation and capitalization companion for the conversation training split of CoRal-project/coral-v3. It contains 102,226 restored transcript rows and no audio bytes.

Pinned companion revision: 6e4fbafde87fbffadd58bbe39a3a2e09e884351a. SHA-256 of data/train-00000-of-00001.parquet: 79d30671f238e884a5b71b682bc811043e7df075566ce5565a236e66ffe32000.

Each row can be joined back to the gated source dataset:

• id_conversation: source conversation utterance identifier
• row_index: source-row position used during generation
• split: source split, currently train
• text: restored punctuation and capitalization
• duration: copied duration when available
• speaker_id: copied speaker identifier when available

Creation process

1. Start from the final ASR manifest rows aligned to CoRal conversation train.
2. Rejoin rows to the source dataset using the retained row-index/identifier cache.
3. Group neighboring utterances as conversational context.
4. Restore Danish punctuation and capitalization over each context window.
5. Write the restored text back to each original utterance without inserting, deleting, or reordering words.
6. Publish only text, join keys, and copied metadata.

The same local coral_conv_pnc_train.jsonl source was sampled directly by Ekko v1's final PnC-v2 training stage. The Hub repository is a joinable companion for that supervision, not a replacement for the source speech corpus.

Intended use

• Punctuation-aware Danish ASR training
• Danish punctuation and capitalization restoration
• Reconstructing the text targets used by Ekko v1 PnC v2 when combined with authorized access to the original CoRal audio

Limitations

• Punctuation and casing were restored automatically, not manually relabeled.
• Context windows improve sentence boundaries but do not make them authoritative ground truth for archival, legal, or accessibility use.
• row_index joins require a compatible source-dataset revision; prefer id_conversation when available.
• This companion currently covers CoRal conversation train only, not read-aloud or held-out evaluation splits.
• No audio is included. Access to source audio remains governed by the original gated CoRal repository.
• An automated pattern screen found no email addresses, URLs, or CPR-like identifiers, but flagged 16 phone-like candidates. The restoration process enforces source word-token parity, and the release review retained those source-token candidates without redaction; see AUDIT.md.

License and access

The source corpus is distributed under CoRal's modified OpenRAIL-D terms. Use of this derived text and any join to source audio must comply with the original CoRal license and access conditions. A complete copy is included as LICENSE, and NOTICE records the transformation. This repository uses gated public access so downstream users acknowledge and follow the same use restrictions as the source dataset.

Source dataset reference

Original dataset:

• CoRal-project/coral-v3, conversation configuration
• Hugging Face: https://huggingface.co/datasets/CoRal-project/coral-v3

Use this companion together with the citation requested by the original CoRal dataset:

@dataset{coral2024,
  author    = {Dan Saattrup Smart, Sif Bernstorff Lehmann, Simon Rands Leminen, Anders Jess Pedersen, Anna Katrine van Zee and Torben Blach},
  title     = {CoRal: A Diverse Danish ASR Dataset Covering Dialects, Accents, Genders, and Age Groups},
  year      = {2024},
  url       = {https://hf.co/datasets/alexandrainst/coral},
}

Also identify the derived target set when relevant:

@dataset{ryeai_coral_v3_conversation_pnc_da,
  title = {CoRal v3 Conversation PnC DA},
  author = {RyeAI},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/RyeAI/coral-v3-conversation-pnc-da}
}