Dataset card for CAESAR-TV3

Dataset Summary

This corpus includes 5 hours and 45 minutes of Catalan speech code-switched with Spanish extracted from the original tv3_parla dataset.

Supported Tasks and Leaderboards

The CAESAR-TV3 dataset is designed for the Automatic Speech Recognition (ASR) task, enabling the transcription of utterances in Catalan, Spanish, and code-switched speech between the two languages.

Languages

The dataset features code-switched speech, combining Catalan (ca) and Spanish (es) within the same audio samples.

Dataset Structure

Data Instances

{
'audio':
  {
    'path': '1429389_1303379885477_289.900_296.740.wav',
    'array': array([0.04263306, 0.06085205, 0.0710144 , ..., 0.04855347, 0.05911255,
       0.03530884]),
    'sampling_rate': 16000
  },
'transcription': "els dies de tempesta les onades fan un so esgarrifós en l'angosta fenedura de sa roncadora"
}

Data Fields

• audio (dict): A dictionary containing the path to the downloaded audio file, the decoded audio array, and the sampling rate.
• text (str): Transcription of the audio file.

Data Splits

The dataset is split into "train", "validation", and "test".

Data loading

from datasets import DownloadConfig, load_dataset
data = load_dataset("BSC-LT/CAESAR-TV3", download_config=download_config, data_dir="data")

Dataset Creation

The original set was created by Baybars Külebi and Alp Öktem from Collectivat. However, the selection and curation of the audios containing ca-es code-switched data was made by Jacobo Romero-Diaz.

Curation Rationale

This corpus specifically focuses on Catalan code-switched with Spanish, a linguistic phenomenon that is very common in the daily lives of Catalonians. This task is particularly low-resourced because, besides being a variety of the Catalan language, it further restricts the available data by incorporating code-switching, a complex and less-explored aspect of language use.

Source Data

This corpus was extracted from the original tv3_parla dataset that includes 240 hours of Catalan speech from broadcast material.

Data Collection and Processing

To extract the CS part, we used the BERT detection. Google’s multilingual BERT was fine-tuned for token classification using a synthetic corpus of code-switched dialogues in Catalan and Spanish. During fine-tuning, each word was labeled with its corresponding language token. Once trained, the model was applied to the transcriptions of the original TV3 Parla dataset, where it performed token-level language classification. This process resulted in a "language count" for each audio file, indicating the distribution of Catalan and Spanish within the transcription. Given that the audios were short, the audio was considered code-switched if Catalan and Spanish were present with at least three words each. With this method, we identified a substantial portion of code-switched data, totaling approximately 5 hours and 45 minutes.

Annotations

The dataset doesn't contain any additional annotations.

Personal and Sensitive Information

The dataset consists of speech from broadcast material. You agree not to attempt to determine the identity of speakers in this dataset.

Considerations for Using the Data

Social Impact of Dataset

CAESAR-TV3 is a source of spontaneous Code-switching speech data that will be valuable in the development of speech technologies for Catalan.

Discussion of Biases

No specific bias mitigation strategies were applied to this dataset. Inherent biases may exist within the data.

Other Known Limitations

Speakers, their gender, and age are not identified, and one or more speakers could be speaking in the same recording. For these reasons, we don't know the total number of speakers in the corpus and their gender/age.

Dataset Curators

The corpus was curated by Jacobo Romero-Diaz in 2024 at the Barcelona Supercomputing Center.

Licensing Information

Creative Commons Attribution Non-Commercial 4.0

Citation Information

@misc{caesar-tv3-bsc2025,
        title={CAESAR collection for Catalan and Spanish Code-switching datasets}, 
        author={Romero-Diaz, Jacobo and Messaoudi, Abir and Armentaro, Carme and Giraldo, José},
        publisher={Barcelona Supercomputing Center},
        year={2025},
        url={https://huggingface.co/datasets/BSC-LT/CAESAR-TV3}
}

Contributions

This work has been promoted and financed by the Generalitat de Catalunya through the Aina project.