VoxLingua107: a Dataset for Spoken Language Recognition

TL;DR

This paper presents VoxLingua107, a large, publicly available dataset for spoken language recognition created from automatically retrieved YouTube videos across 107 languages, demonstrating competitive performance with proprietary datasets.

Contribution

The paper introduces VoxLingua107, a novel large-scale, automatically collected speech dataset for 107 languages, with high labeling accuracy, enabling effective spoken language recognition research.

Findings

Automatic data collection yields high-quality labeled data.

Models trained on VoxLingua107 perform competitively with proprietary datasets.

The dataset supports multiple spoken language identification tasks.

Abstract

This paper investigates the use of automatically collected web audio data for the task of spoken language recognition. We generate semi-random search phrases from language-specific Wikipedia data that are then used to retrieve videos from YouTube for 107 languages. Speech activity detection and speaker diarization are used to extract segments from the videos that contain speech. Post-filtering is used to remove segments from the database that are likely not in the given language, increasing the proportion of correctly labeled segments to 98%, based on crowd-sourced verification. The size of the resulting training set (VoxLingua107) is 6628 hours (62 hours per language on the average) and it is accompanied by an evaluation set of 1609 verified utterances. We use the data to build language recognition models for several spoken language identification tasks. Experiments show that using the automatically retrieved training data gives competitive results to using hand-labeled proprietary datasets. The dataset is publicly available.