palli23/Spjallromur-AB-Diarization-test

Dataset Card — Spjallrómur AB Diarization Test Set

Dataset Description

Spjallrómur AB Diarization Test Set is a curated test set for speaker diarization evaluation on conversational Icelandic speech. It is derived from the Spjallrómur 26.03 corpus by merging the two separate recording channels (A and B) into single mixed-audio files, producing a realistic diarization benchmark with known ground-truth speaker turns.

The test set was compiled as part of the Almannarómur project: Towards Automatic Speech Recognition for Conversations in Icelandic (Project code: MVF25010163).

• Curated by: Páll Rúnarsson
• Funded by: Almannarómur (Language Technology Programme for Icelandic)
• Language(s): Icelandic (is)
• License: CC BY-SA 4.0 — free to use, share, and adapt, provided you give appropriate credit and distribute any derivatives under the same license.
• Source corpus: Spjallrómur 26.03
• Repository: palli23/Spjallromur-AB-Diarization-test

Dataset Sources

• Source corpus: Spjallrómur 26.03 — CLARIN-IS
• Project: Almannarómur

---

Uses

Direct Use

This dataset is intended for evaluation and benchmarking of speaker diarization systems on conversational Icelandic speech. By merging independently recorded speaker channels into a single mixed audio stream, the dataset provides realistic overlapping and adjacent speech conditions with exact ground-truth speaker labels and turn boundaries.

The primary metric for evaluation on this dataset is Diarization Error Rate (DER).

Out-of-Scope Use

• Training or fine-tuning diarization models (this is a test-only split)
• Languages other than Icelandic
• ASR-only evaluation without speaker labelling (use palli23/spjallromur-4h instead)

---

Dataset Structure

This dataset contains a single test split only. It should not be used for training or hyperparameter tuning, as it is intended solely for final evaluation.

Split	Content
test	Mixed A+B channel audio + reference speaker annotations

Data Fields

• audio — merged A+B channel audio (16 kHz, mono mix)
• rttm (or equivalent) — ground-truth speaker turn annotations with timestamps and speaker labels
• (add any additional fields)

Data Collection and Processing

The source recordings in Spjallrómur were captured with each conversation participant on a separate channel (A and B), enabling clean per-speaker audio. For this dataset, the two channels are mixed into a single audio stream to simulate a realistic single-microphone diarization scenario. Ground-truth speaker turn boundaries and labels are derived from the manual transcriptions included in Spjallrómur 26.03.

---

Dataset Creation

Curation Rationale

Speaker diarization benchmarks for Icelandic are essentially non-existent. This dataset was compiled to provide a standardised, reproducible test set for diarization evaluation on Icelandic conversational speech, enabling fair comparison of diarization systems within the Almannarómur project and beyond.

The channel-merge approach leverages the clean per-speaker recordings in Spjallrómur to produce ground truth of known quality, while the mixed output reflects realistic single-channel conditions encountered in real deployments.

Source Data

Derived from Spjallrómur 26.03, compiled and processed by Páll Rúnarsson. Channel merging and annotation alignment were performed as part of the Almannarómur project.

Annotations

Ground-truth speaker turn annotations are derived from the manual transcriptions in Spjallrómur 26.03, produced by three transcribers and including turn-level timestamps and speaker labels.

---

Bias, Risks, and Limitations

• Speaker demographics reflect those of the Spjallrómur corpus; results may not generalise to all Icelandic speaker populations.
• The channel-merge process produces a controlled overlap scenario; real-world microphone conditions may introduce additional noise and reverberation not captured here.
• This dataset covers spontaneous conversational Icelandic only; diarization performance on other speech styles (broadcast, read speech) is out of scope.

Recommendations

When reporting DER on this test set, users should clearly specify:

• Whether collar tolerance is applied (and its duration)
• How overlapping speech regions are handled
• The diarization system's use of oracle vs. predicted number of speakers

---

Citation

If you use this dataset, please cite both this work and the source corpus:

@misc{runarsson2026spjallromur_ab,
  author       = {Rúnarsson, Páll},
  title        = {Spjallrómur AB -- Icelandic Conversational Diarization Test Set},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/palli23/Spjallromur-AB-Diarization-test}}
}

@misc{fong2026spjallromur,
  author       = {Fong, Judy Y. and Borsky, Michal and Runarsson, Pall
                  and Hedström, Staffan and Jónsson, Ólafur Helgi
                  and Hólmfriðardóttir, Lára Margrét H. and Þorsteinsdóttir, Sunneva
                  and Eiríksdóttir, Málfriður Anna and Mollberg, David Erik
                  and Magnúsdóttir, Eydís Huld and Þórhallsdóttir, Ragnheiður
                  and Gudnason, Jon},
  title        = {Spjallromur 26.03 -- Icelandic Conversational Speech},
  year         = {2026},
  publisher    = {CLARIN-IS / Reykjavík University},
  howpublished = {\url{http://hdl.handle.net/20.500.12537/379}}
}

---

Acknowledgements

This dataset was compiled at the Language and Voice Laboratory (lvl.ru.is) at Reykjavík University, Iceland.

Funded by Almannarómur — the Language Technology Programme for Icelandic, as part of the project Towards Automatic Speech Recognition for Conversations in Icelandic (MVF25010163), funded by the Icelandic Ministry of Education, Science and Culture.

---

Dataset Card Author

Páll Rúnarsson, Language and Voice Laboratory, Reykjavík University

Dataset Card Contact

(add contact email or HF profile link)