VOXLINGUA107: A Dataset for Spoken Language Recognition

By Jörgen Valk, Tanel Alumäe. 2020

234 citations recorded by the source corpus.

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 available1.

Subfields: Artificial Intelligence, Evaluation & Benchmarks, Speech & Audio, Signal Processing

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Evaluation & Benchmarks

In Speech & Audio

In Signal Processing

Citations in the corpus

Robust Speech Recognition via Large-Scale Weak Supervision (2022, methodology): “We also use an audio language detector, which was created by fine-tuning a model trained on a prototype version of the dataset on VoxLingua107 (Valk & Alum¨ae, 2021) to ensure that the spoken language matches the language of the transcript according to CLD2.”
FLEURS: FEW-Shot Learning Evaluation of Universal Representations of Speech (2022, background): “9k hrs Open domain Read Yes Yes No VoxPopuli [14] 24 400k hrs* Parliament Spontaneous Partial Partial Partial BABEL [17] 25 2k hrs Conversational Spontaneous Yes No No CommonVoice [16] 93 15k hrs Open domain Read Yes No No Voxlingua-107 [22] 107 6.”
MAESTRO: Matched Speech Text Representations through Modality Matching (2022, methodology): “(XLS-R uses 8k extra hours of untranscribed speech from VoxLingua [39].”
XLS-R: Self-supervised Cross-lingual Speech Representation Learning at Scale (2021, methodology): “We consider VoxLingua107 [23] which spans 107 languages.”
XTREME-S: Evaluating Cross-lingual Speech Representations (2022, methodology): “LangID can be evaluated using VoxLingua107 [25] on YouTube data in 107 languages, and CMU Wilderness [26] on New Testament data in 700+ languages.”
Self-Supervised Speech Representation Learning: A Review (2022, unknown): “Several datasets, including GigaSpeech [220], TED-LIUM 3 (TED3) [221], TED-LIUM 2 (TED2) [222], Switchboard (SWB) [223], TIMIT [96], and VoxLingua107 [224], are labeled”
Automatic speaker verification spoofing and deepfake detection using wav2vec 2.0 and data augmentation (2022, unknown): “0 XLS-R is a large-scale cross-lingually pre-trained model trained on diverse corpora including VoxPopuli data (VP-400K) [41], the multilingual Librispeech corpus (MLS) [42], CommonVoice (CV) [43], VoxLingua107 (VL) [44], and BABEL (BBL) [45] datasets.”
Scaling Speech Technology to 1, 000+ Languages (2023, methodology): “Training a language identification system requires speech data for which the spoken language is known, however, the most diverse public corpora span no more than about 100 languages [Valk and Alumäe, 2020, Conneau et al., 2022].”