VoxLingua107: 107-Language Speech Dataset for Language ID

VoxLingua107 is a speech dataset for training spoken language identification models. It contains 6628 hours of short speech segments sourced from YouTube videos, covering 107 languages. The dataset was created by SEACrowd and was last updated in June 2024.

Use Cases

• Train language identification models based on short speech segments.
• Benchmark audio classification algorithms on a multilingual dataset.
• Develop speech recognition systems for underrepresented languages based on the 107-language coverage.
• Validate language detection accuracy using the volunteer-verified development set.

Strengths

• Covers 107 languages, providing broad linguistic diversity.
• Contains 6628 total hours of speech data.
• Includes a separate development set of 1609 segments validated by at least two volunteers.

Limitations

• The amount of data per language varies significantly, averaging 62 hours.
• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.

Provenance

Source

YouTube videos.

Collection Method

Speech segments sourced from YouTube videos, labeled based on language indicated in video title and description.

Freshness

Last updated 2024-06-24 13:32:12; freshness should be verified.