Omnilingual ASR Corpus: Spontaneous Speech for 348 Under-Served Languages

348 under-served languages are represented in this collection of spontaneous speech recordings and transcriptions. The corpus was collected by Meta FAIR's Omnilingual ASR project for training automatic speech recognition and spoken language identification models. It was last updated on the Hugging Face platform in December 2025.

Use Cases

• Training automatic speech recognition (ASR) models based on the described spontaneous speech recordings.
• Training spoken language identification models based on the multilingual audio data.
• Benchmarking speech technology performance across a wide range of under-served languages.
• Studying linguistic features and acoustic patterns in low-resource language communities.

Strengths

• Covers a large number of 348 under-served languages.
• Specifically designed for training ASR and language ID models, as stated in the description.
• Includes both audio recordings and transcriptions.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

Meta FAIR's Omnilingual ASR project.

Collection Method

Collected as part of a project for training ASR and language ID models.

Freshness

Last updated 2025-12-18 19:43:25.