distilled-catalan-youtube-speech

# Dataset Card for distilled-catalan-youtube-speech ## Table of Contents - [Dataset Description](#dataset-description) - [Dataset Summary](#dataset-summary) - [Supported Tasks](#supported-tasks-and-leaderboards) - [Languages](#languages) - [Dataset Structure](#dataset-structure) - [Data Instances](#data-instances) - [Data Fields](#data-fields) - [Data Splits](#data-splits) - [Dataset Creation](#dataset-creation) - [Curation Rationale](#curation-rationale)This work has been promoted and financed by the Generalitat de Catalunya through the [Aina project](https://projecteaina.cat/). - [Source Data](#source-data) - [Annotations](#annotations) - [Personal and Sensitive Information](#personal-and-sensitive-information) - [Considerations for Using the Data](#considerations-for-using-the-data) - [Social Impact of Dataset](#social-impact-of-dataset) - [Discussion of Biases](#discussion-of-biases) - [Other Known Limitations](#other-known-limitations) - [Additional Information](#additional-information) - [Dataset Curators](#dataset-curators) - [Licensing Information](#licensing-information) - [Citation Information](#citation-information) - [Funding](#funding) ## Dataset Description - **Homepage:** [Project Aina](https://www.bsc.es/research-and-development/not-assigned-pages/about-aina) - **Repository:** [Distilled Catalan YouTube Speech Corpus](https://huggingface.co/datasets/langtech-veu/distilled-catalan-youtube-speech) - **Point of Contact:** [Language Technologies Unit]([e-mail]) ### Dataset Summary The [Distilled Catalan YouTube Speech](https://huggingface.co/datasets/langtech-veu/distilled-catalan-youtube-speech) Corpus is a curated subset of the [Catalan YouTube Speech](https://huggingface.co/datasets/softcatala/catalan-youtube-speech) Corpus by [Softcatalà](https://huggingface.co/softcatala), enriched with automatically verified transcriptions generated using two independent ASR systems known as verification models. These models were trained on different data sources, and their agreement is used as a proxy for transcription reliability. The dataset includes 207 hours of transcribed Catalan speech, with transcriptions categorized by quality: perfect matches (identical outputs between systems) and word count matches (same word count but different wording, resolved using a third ASR system). In addition, a manually annotated test set and a high-confidence validation set are provided. This corpus offers a reliable source of spontaneous Catalan speech and is intended to support research and development in automatic speech recognition (ASR), especially in low-resource and semi-supervised settings. ### Example Usage The Distilled Catalan YouTube Speech Corpus is divided into 4 loadable splits: perfect_matches, word_count_matches, validation, and test. To load the whole dataset, do: ```python from datasets import load_dataset ds_dcys = load_dataset("langtech-veu/distilled-catalan-youtube-speech") ``` To load a specific split (for example, the split with the best quality transcripts), do: ```python from datasets import load_dataset ds_dcys_pm = load_dataset("langtech-veu/distilled-catalan-youtube-speech",split="perfect_matches") ``` ### Supported Tasks automatic-speech-recognition: The dataset can be used to train a model for Automatic Speech Recognition (ASR). The model is presented with an audio file and asked to transcribe the audio file to written text. The most common evaluation metric is the word error rate (WER). ### Languages The audio is in Catalan. ## Dataset Structure ### Data Instances ```python { 'audio_id': 'fdb498ff-3745-4615-b676-c8cc11206939', 'audio': { 'path': '/home/carlos/.cache/HuggingFace/datasets/downloads/extracted/0b7d329065bdd7754db2d5938929a50e2c441b15c2b809b70dd19f7c3550938a/pm_fem_part_001/a1QL6wElEAQ/d81eebdf-5146-4b0b-a13f-2c3292dbd9ca.flac', 'array': array([0.0007019 , 0.00039673, 0.00064087, ..., 0.0012207 , 0.00146484, 0.0017395 ]), 'sampling_rate': 16000 }, 'corpus_id': 'distilled-catalan-youtube-speech', 'split': 'train-perfect_matches', 'language': 'Catalan', 'duration': 6.5279998779296875, 'gender': 'female', 'yt_url': 'aOp48D2AXw0&t=3743', 'consensus': 'AB', 'selected_trans': 'A', 'normalized_text': 'a vegades maltractant francament molt el la base arquitectònica' } ``` ### 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. * `corpus_id` (string) - id of the dataset * `split` (string) - Indicates both the dataset split (train, validation, or test) and the transcription quality category (human_annotated, perfect_matches, or word_count_matches). * `language` (string) - Language of the speech segment. In this dataset, all segments are in Catalan. * `duration` (float32) - Duration of the audio file in seconds. * `gender` (string) - Speaker gender (male or female), as provided by the original [Catalan YouTube Speech](https://huggingface.co/datasets/softcatala/catalan-youtube-speech) corpus. * `yt_url` (string) - YouTube video identifier. When appended to https://www.youtube.com/watch?v=, it leads to the original video. The segment corresponds to a specific time interval within that video. * `consensus` (string) - Indicates which systems agreed on the number of words. For instance, A1[33] means that Model A and Candidate 1 both generated transcriptions with 33 words. * `selected_trans` (string) - In the word_count_matches category, even though systems agree on word count, their transcriptions may differ lexically. This field specifies which of the two transcriptions was selected using a third ASR system (different from Models A and B). * `normalized_text` (string) - Final transcription after normalization (e.g., lowercasing, punctuation removal, etc.). ### Data Splits The corpus is divided into 2 categories based on the quality of the transcriptions. * **perfect_matches** (131h48m12s | 41,865 files): Segments for which one or more verification models produced identical transcriptions, either with each other or with one of the candidate transcriptions. * **word_count_matches** (70h47m43s | 15,621 files): Segments where one or more verification models agreed only on the number of words in the transcriptions, regardless of lexical differences. To determine which of the two transcriptions to retain, a third ASR model (different from Models A and B) is used. The transcription with the lowest Word Error Rate (WER) relative to the third model is selected. In this split, the Character Error Rate (CER) between the two candidate transcriptions is ≤ 2.63%, and we estimate the overall WER to be ≤ 4%. The corpus also has validation (dev) and test splits: * **test** (02h56m02s | 711 files): Manually annotated speech recordings used for evaluation. * **validation** (02h20m45s | 1,456 files): A high-quality subset of the perfect_matches category, where both verification models (A and B) and both candidate transcriptions (1 and 2) are in complete agreement. To load a specific portion, please check the above section "Example Usage". ## Dataset Creation ### Curation Rationale The motivation for curating this dataset stems from the need for high-quality ASR training data in Catalan, a language with limited available resources. While the original [Catalan YouTube Speech](https://huggingface.co/datasets/softcatala/catalan-youtube-speech) corpus provides large quantities of speech, its transcriptions vary in quality. To distill the most reliable segments, we trained two independent ASR systems (verification models [A](https://huggingface.co/langtech-veu/catalan-verification-model-pkt-a) and [B](https://huggingface.co/langtech-veu/catalan-verification-model-pkt-b)) and selected transcriptions based on system agreement. Perfect agreement was used as a strong indicator of correctness, while in cases of partial agreement (word count matches), a third ASR system was used to select the more accurate hypothesis. This approach enables the creation of a high-confidence dataset with minimal human effort, making it especially valuable for training robust ASR models in under-resourced languages. ### Source Data #### Initial Data Collection and Normalization The audio data in this corpus was sourced directly from the [Catalan YouTube Speech](https://huggingface.co/datasets/softcatala/catalan-youtube-speech) dataset developed by [Softcatalà](https://huggingface.co/softcatala). No additional recordings were collected. We did not alter the original segmentation or reprocess the audio files. Instead, we focused on curating the transcriptions by applying an automatic verification strategy. Specifically, the corpus was processed using Verification Models [A](https://huggingface.co/langtech-veu/catalan-verification-model-pkt-a) and [B](https://huggingface.co/langtech-veu/catalan-verification-model-pkt-b), trained independently on different datasets. Segments were retained based on model agreement—either through identical transcriptions or matching word counts, with final selection resolved via a third ASR system. This process produced a high-confidence subset of the original corpus. ### Annotations #### Annotation process Only the test split was manually annotated. The rest was automatically verified through verification models [A](https://huggingface.co/langtech-veu/catalan-verification-model-pkt-a) and [B](https://huggingface.co/langtech-veu/catalan-verification-model-pkt-b). #### Who are the annotators? The annotators are part of the staff of the [Barcelona Supercomputing Center](https://www.bsc.es/) and they were led by [Carme Armentano-Oller](https://huggingface.co/carmentano). ### Personal and Sensitive Information The dataset consists of public YouTube videos with a CC license. You agree not to attempt to determine the identity of speakers in this dataset. ## Considerations for Using the Data ### Social Impact of Dataset The Distilled Catalan YouTube Speech Corpus is a source of spontaneous speech data that will be valuable in the development of speech technologies for Catalan. ### Discussion of Biases The language is limited to the YouTube videos used to create the corpus and may not be representative of all domains. ### Other Known Limitations * Speaker gender information was not determined by us, but rather inherited from the original [Catalan YouTube Speech](https://huggingface.co/datasets/softcatala/catalan-youtube-speech) corpus. We have not verified the accuracy or consistency of this metadata. * 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. * Code-switching detection was not applied. Some segments may include speech in other languages (e.g., Spanish or English), but no annotation or estimation of code-switching frequency or distribution has been conducted. ## Additional Information ### Dataset Curators The corpus was curated by [Carlos Daniel Hernández Mena](https://huggingface.co/carlosdanielhernandezmena) in 2025 in the [Language Technologies Laboratory](https://huggingface.co/BSC-LT) of the [Barcelona Supercomputing Center](https://www.bsc.es/). The collection of the manual annotations that constitute the test split was performed by [Carme Armentano-Oller](https://huggingface.co/carmentano). ### Licensing Information [MIT](https://opensource.org/licenses/MIT) ### Citation Information ``` @misc{BSC-discatyouspe-2025, title={The Distilled Catalan YouTube Speech Corpus}, author={Hern{\'a}ndez Mena, Carlos Daniel and Armentano-Oller, Carme}, organization={Barcelona Supercomputing Center}, year={2025} } ``` ### Contact For further information, please send an email to <[e-mail]>. ### Funding This work/research has been promoted and financed by the Government of Catalonia through the [Aina project](https://projecteaina.cat/). The training of the model was possible thanks to the computing time provided by [Barcelona Supercomputing Center](https://www.bsc.es/) through MareNostrum 5. We acknowledge the EuroHPC Joint Undertaking for awarding us access to MareNostrum 5.