Ugiat/voxlingua107_IberLang

VoxLingua107 IberLang Dataset

Dataset Summary

The IberVoice dataset is a curated and validated collection of audio samples in Spanish, Catalan, Galician, Basque and Occitan designed to enhance spoken language identification (LID) systems for the official languages of Spain. The dataset originates from the VoxLingua107 multilingual corpus, but has been cleaned and reannotated to correct systematic labeling errors affecting minority languages.

Why

In the original VoxLingua107 dataset, language labels were assigned automatically based on YouTube video metadata, resulting in numerous mislabeled samples — particularly for Catalan, Galician, Basque and Occitan where Spanish speech was often included under incorrect tags.

Correction Process

To address this, each subset of IberVoice (except Occitan) underwent a two-stage refinement process:

1. High-confidence filtering: Audio segments with a language confidence score above 0.95, as predicted by the Whisper Large model, were retained to ensure linguistic purity.

2. Iterative reclassification: A Whisper Medium model was fine-tuned on these clean samples for language identification. The resulting model was then used to predict labels for the remaining audios. Samples with prediction confidence above 0.8 were automatically reassigned, while low-confidence cases were manually verified.

This process produced a high-quality, balanced, and linguistically faithful dataset for the four Iberian languages.

For Occitan, since there was less data to check, a native speaker reviewed each of the audio files and discarded any that weren't well-labeled.

If you want to see all the changes made to the original dataset, check the file Errors_found.ods.

Language Coverage

Language	Code	Hours
Catalán	ca	83
Gallego	gl	58
Español	es	39
Euskera/Basque	eu	27
Occitan	oc	13

Citation

This dataset is derived from VoxLingua107 by J. Valk and T. Alumae (2021), licensed under CC BY 4.0.

Modifications: dataset cleaning and language reclassification for Catalan, Galician, Basque, Spanish, and Occitan, performed by Ugiat Technologies.