YapayNet/cs-fleurs-tagged

CS-FLEURS Tagged

CS-FLEURS Tagged provides tagged transcripts for test split of Russian-English and German-English code-switching utterances. The annotations mark embedded English words in the reference transcript so they can be evaluated directly with Point-of-Interest Error Rate (PIER).

Tagged words use the following format:

<tag WORD>

For example:

<tag Romanticism> во многом включал в себя <tag cultural> детерминизм

In this example, Romanticism and cultural are the embedded English words of interest.

Dataset Contents

The dataset currently contains two CS-FLEURS-test language pairs:

Language pair	Description
rus-eng	Russian matrix language with embedded English words
deu-eng	German matrix language with embedded English words

Each example includes the utterance audio, a tagged transcript, a plain transcript with tags removed, and the list of tagged embedded English words.

Columns

• audio: utterance audio
• id: utterance id
• language_pair: language pair, for example rus-eng or deu-eng
• matrix_language: matrix language code
• embedded_language: embedded language code, currently eng
• duration_ms: utterance duration in milliseconds
• text_tagged: transcript with <tag ...> annotations
• text_plain: transcript with tags removed
• embedded_words: list of tagged embedded English words
• source_dataset: original source dataset id

Usage

Load the dataset with datasets:

from datasets import load_dataset

ds = load_dataset("YapayNet/cs-fleurs-tagged", split="test")

example = ds[0]
print(example["id"])
print(example["text_tagged"])
print(example["embedded_words"])
print(example["audio"])

The text_tagged field can be used when the evaluation needs to preserve the explicit points of interest. The text_plain field is useful for standard ASR evaluation or display. The embedded_words field provides the tagged words as a list.

PIER Evaluation

This dataset is designed for code-switching ASR evaluation with PIER, a metric that focuses evaluation on selected words of interest instead of only reporting aggregate WER.

Use the PIER evaluation code from:

enesyugan/PIER-CodeSwitching-Evaluation

The annotation and tagging setup is described in:

enesyugan/robust-code-switching-asr

Source Dataset

This dataset is derived from CS-FLEURS test split:

byan/cs-fleurs

Audio and base transcripts come from CS-FLEURS. The <tag ...> annotations were added for PIER-style code-switching evaluation.

License

This dataset follows the license of the source dataset: CC BY-NC 4.0.

Citation

If you use this dataset, please cite CS-FLEURS, how it was tagged ugan2026adding and if you evaluate using PIER ugan2025pier.

@article{ugan2026adding,
  title={Adding Robust Code-Switching Capabilities to High Performance Multilingual ASR},
  author={Ugan, Enes Yavuz and Waibel, Alexander},
  journal={arXiv preprint arXiv:2606.21990},
  year={2026}
}

@article{yan2025cs,
  title={CS-FLEURS: A Massively Multilingual and Code-Switched Speech Dataset},
  author={Yan, Brian and Hamed, Injy and Shimizu, Shuichiro and Lodagala, Vasista and Chen, William and Iakovenko, Olga and Talafha, Bashar and Hussein, Amir and Polok, Alexander and Chang, Kalvin and others},
  journal={arXiv preprint arXiv:2509.14161},
  year={2025}
}

@misc{ugan2025pier,
  title={PIER: A Novel Metric for Evaluating What Matters in Code-Switching},
  author={Ugan, Enes Yavuz and Pham, Ngoc-Quan and Bärmann, Leonard and Waibel, Alex},
  year={2025},
  eprint={2501.09512},
  archivePrefix={arXiv},
  primaryClass={cs.CL}
}