ASR Datasets
Collection
Datasets for training and evaluation of ASR models.
•
7 items
•
Updated
Datasets:
Please provide the required information and accept the Terms and Conditions to request access to the dataset
Terms and Conditions
This dataset may be used solely for research purposes and for developing artificial intelligence tools focused on linguistic objectives. It can be accessed by any entity collaborating on projects promoted by the Centro Ramón Piñeiro de Investigación en Humanidades and Universidade de Santiago de Compostela, particularly those involved in the implementation of the Nós project. Dissemination of the voice recordings in an open-access manner or their public exposure is strictly prohibited. By accessing and using this dataset, you agree to comply with all applicable laws and ethical standards regarding the protection of individual rights. Users are strictly prohibited from using the dataset in any way that infringes upon the rights, privacy, or dignity of any individual represented within it. Any misuse, including but not limited to attempts to engage in discriminatory, harmful, or unlawful activities, is expressly forbidden.
This document is a legal contract between you and the dataset provider. By accessing and using the dataset, you acknowledge and agree to abide by the Terms and Conditions.
Our team may take a few days to process your access request. Thank you in advance for your patience.
Log in or Sign Up to review the conditions and access this dataset content.
Corpus description
Nos_Telexornal-GL is an ASR corpus of more than 1,100 hours of automatically transcribed and aligned speech using Google Chirp2. This corpus was created from audio of Galician news between 2019 and 2022. The content belongs to the Galician Public Television and Radio Corporation (public broadcaster for the autonomous community of Galicia) and the data is released according to their terms of use.
The corpus is split into two subcorpora, clean and other. To ensure high-quality transcriptions, the Chirp2 transcriptions where compared with Mozilla's Whisper-large-v3-gl, and only those of WER lower than 15% between models where selected for the clean subcorpus, whereas the other subcorpus comprises the segments that were discarded in the filtering process. A manual review was also conducted to correct the most serious errors. The details of both subcorpora can be found in the table below:
| Subcorpus | No. of hours | No. of segments |
|---|---|---|
| clean | 662.17 | 87,336 |
| other | 509.93 | 60,186 |
| Total | 1,172.1 | 147,522 |
For each subcorpora, we provide three splits: train, dev and test. Test and dev splits contain 15 hours of audio.
The repository structure is as follows: the root folder contains all the segments, divided into clean and other, and also split into train, dev and test for each subcorpora. Each segment has the following fields:
filename: the name of the audio file.audio: audio file.sentence: the transcript of the audio file.
The other subcorpus contains some audio files that are several minutes long and others that are too short, so a filtering process may be necessary before training.
Funding and acknowledgements
This dataset was compiled within the Nós Project, funded by the Ministerio para la Transformación Digital y de la Función Pública - Funded by EU – NextGenerationEU within the framework of the project ILENIA with reference 2022/TL22/00215336. It also received support from the Investigo program (RTRP - NextGenerationEU).
We would like to thank the Galician Public Television and Radio Corporation (CRTVG) for their kind collaboration in providing the original data.
For more information, please go to nos.gal or contact the Nós project at [email protected].
Licensing Information
Creative Commons Attribution 4.0 International
Terms and Conditions
By accessing and using this dataset, you agree to comply with all applicable laws and ethical standards regarding the protection of individual rights. The dataset contains voice files, transcripts, and metadata, including participant identity information, provided solely for research and development purposes. Users are strictly prohibited from using the dataset in any way that infringes upon the rights, privacy, or dignity of any individual represented within it. Any misuse, including but not limited to attempts to engage in discriminatory, harmful, or unlawful activities, is expressly forbidden.
Citation information
If you use this version of the dataset, please cite as follows:
Antonio Moscoso Sánchez, Carmen Magariños, Daniel Fernández López, María Pérez Lago, Adina Ioana Vladu, Natalia Villar Martínez, Carla Castedo, Francisco Dubert-García, Xosé Luis Regueira Fernández & Elisa Fernández Rei. (2025). Nos_Telexornal-GL. URL: https://huggingface.co/datasets/proxectonos/Nos_Telexornais-GL
- Downloads last month
- 7