BSC-LT/BSCs_Code_Switching_CA-ES_ASR_Test

Dataset Card for distilled-catalan-youtube-speech

Dataset Summary

BSC's Code-Switching Catalan-Spanish ASR Test is a manually curated evaluation set designed to assess automatic speech recognition (ASR) systems in a code-switching scenario involving Catalan and Spanish. The corpus contains 867 audio recordings totaling 4 hours and 9 minutes, where Catalan is the dominant language, but segments in Spanish are systematically present within each utterance.

The recordings were extracted from two existing public corpora: Corts Valencianes Anonymized and Parlament Parla Version 3. Each utterance was manually inspected to ensure the presence of intra-sentential code-switching, making this dataset a reliable benchmark for evaluating ASR models under realistic bilingual speech conditions.

Example Usage

The BSC's Code-Switching Catalan-Spanish ASR Test contains only the test split. To load the whole dataset, do:

from datasets import load_dataset
ds_dcys = load_dataset("BSC-LT/BSCs_Code_Switching_CA-ES_ASR_Test")

To load the test split do:

from datasets import load_dataset
ds_dcys_pm = load_dataset("BSC-LT/BSCs_Code_Switching_CA-ES_ASR_Test",split="test")

Supported Tasks

automatic-speech-recognition: The dataset can be used to test a model for Automatic Speech Recognition (ASR). The model is presented with an audio file of this dataset and asked to transcribe it to written text. The most common evaluation metric is the word error rate (WER).

Languages

The dataset features Catalan speech with embedded Spanish segments due to code-switching.

Dataset Structure

Data Instances

{
  'audio_id': '00dce5ad7875a60c7ffaa6609c20c98e', 
  'audio': {
    'path': '/home/carlos/.cache/HuggingFace/datasets/downloads/extracted/4dc6db5456b2e885f92476a633ea8d04cf52dc997bb72c2f3cec7ba5edd3c0e5/corts_valencianes_anonymized/other_train_long/0/0/d/00dce5ad7875a60c7ffaa6609c20c98e.flac', 
    'array': array([-1.15016440e-03, -1.61928063e-03, -1.53533490e-03, ...,
        7.47151563e-05, -4.17502355e-04,  0.00000000e+00]), 
    'sampling_rate': 16000
  }, 
  'source': 'corts_valencianes_anonymized', 
  'split': 'test', 
  'duration': 33.599700927734375, 
  'normalized_text': "i jo posant-me a parlar a nivell esquemàtic sobre com ordenaria tot el que ha passat en el brugal i en el pla general d'ordenació urbana primer diria senyora castedo agència de col·locació haurien de canviar el nom de servef per agencia de colocación señora castedo una senyora que es dedica a col·locar a tot aquell que era amic seu o del partit cridant simplement o enviant un missatge al senyor ortiz oiga colóquele en inusa seis meses"
}

Data Fields

• audio_id (string) - Unique identifier for the audio segment.
• audio (datasets.Audio) - A dictionary containing the path to the audio file, the decoded audio array, and the sampling rate. In non-streaming mode (default), the path points to the locally extracted audio file. In streaming mode, it corresponds to the relative path of the audio inside its archive, as files are not extracted locally.
• source (string) - Indicates if the current sample comes from the corpus Corts Valencianes or Parlament Parla
• split (string) - This dataset contains the test split only.
• duration (float32) - Duration of the audio file in seconds.
• normalized_text (string) - Final transcription after normalization (e.g., lowercasing, punctuation removal, etc.).

Data Splits

The corpus contains the test split only with samples taken from Corts Valencianes Anonymized and Parlament Parla Version 3.

Dataset Creation

Curation Rationale

This dataset was curated to address the lack of publicly available evaluation sets for automatic speech recognition (ASR) in code-switching scenarios between Catalan and Spanish. While both languages are widely spoken in bilingual regions of Spain, existing ASR resources treat them independently and do not account for the common phenomenon of speakers alternating between them within a single utterance.

To fill this gap, we created a dedicated test set that captures intra-sentential code-switching in realistic contexts. The recordings were manually selected from two larger public corpora of parliamentary speech. In order to identify utterances containing Spanish segments, we applied a simple yet effective filtering strategy based on lexical cues: we searched for Spanish-specific words such as "y" or "rey", as well as words containing the letter "ñ", which does not exist in standard Catalan. This heuristic helped us efficiently spot potential instances of code-switching, which were then manually verified.

The resulting dataset enables rigorous evaluation of ASR models in Catalan-Spanish bilingual settings, with special focus on real-world code-switching behavior.

Source Data

Initial Data Collection and Normalization

The audio data in this corpus was sourced directly from Corts Valencianes Anonymized and Parlament Parla Version 3.

Annotations

Annotation process

The transcriptions of the source datasets were not altered.

Who are the annotators?

We did not create new annotations for this dataset.

Personal and Sensitive Information

The dataset consists of public audios of parliamentary interventions. You agree not to attempt to determine the identity of speakers in this dataset.

Considerations for Using the Data

Social Impact of Dataset

The BSC's Code-Switching Catalan-Spanish ASR Test is a source of spontaneous speech data that will be valuable in the evaluation of speech technologies for Catalan.

Discussion of Biases

The language is limited to the parliamentary interventions used to create the corpus and may not be representative of all domains.

Other Known Limitations

• The audio samples comming from Corts Valencianes Anonymized were subjected to an anonymization process which means that the audio signal was altered from the original.

• Speaker gender information is not provided.

• No speaker diarization or speaker count verification was performed. As a result, some audio segments may feature multiple speakers.

• Background noise conditions have not been assessed or annotated. The dataset may contain recordings with varying levels of noise, music, or overlapping speech.

Additional Information

Dataset Curators

The corpus was curated in 2025 by the Speech Technologies Team of the Language Technologies Laboratory of the Barcelona Supercomputing Center. The manual selection of the audios was performed by Carlos Daniel Hernández Mena. The experiments for the InterSpeech paper "Optimizing ASR for Catalan-Spanish Code-Switching: A Comparative Analysis of Methodologies" were performed by Pol Serra, Jacobo Romero, Abir Messaoudi, Jose Giraldo, Carme Armentano-Oller, Rodolfo Zevallos and Ivan Meza under the supervision of Javier Hernando, head of the Speech Technologies Team.

Licensing Information

CC-BY-NC-4.0

Citation Information

@misc{mena2025optimizing,
      title={BSC's Code-Switching Catalan-Spanish ASR Test}, 
      author={Hern{\'a}ndez Mena, Carlos Daniel and Serra, Pol and Romero, Jacobo and Messaoudi, Abir and Giraldo, Jose and Armentano-Oller, Carme and Zevallos, Rodolfo and Meza, Ivan and Hernando, Javier},
      organization={Barcelona Supercomputing Center},
      year={2025},
      url={https://huggingface.co/datasets/BSC-LT/BSCs_Code_Switching_CA-ES_ASR_Test},
}

Funding

This work/research has been promoted and financed by the Government of Catalonia through the Aina project.