Basis AI releases expressive-speech dataset spanning 108 locales
Global Emotions contains 525 hours of speech recorded by 10,420 speakers across 83 languages and 108 locales; the company made the dataset available for commercial and research use.
Event / data period:
Published: · Updated:
Basis AI released Global Emotions, an audio dataset designed for controlled comparisons of emotion and speaking style. The company said it contains 525 hours recorded by 10,420 speakers across 108 locales and 83 languages. Participants read the same lines in neutral, medium and full emotional intensities, as well as styles such as whispering and singing.
The company says users must request access through Hugging Face, where requests receive human review; after access is granted, the license allows free commercial and research use. Native speakers from each locale rated the recordings. Audio was not post-processed for denoising, loudness normalization or resampling. The open release does not include human-validated transcripts; the source says those are sold separately and clarifies that `given_line` is the prompt given to a speaker, not a verified transcript.
Data
Scope of the Global Emotions speech dataset
Size of the controlled emotional and speaking-style recordings released by Basis AI.
- Recorded speech hours
- 525 hours
- 2026-10-12 · Observed
- Speakers
- 10,420 count
- 2026-10-12 · Observed
- Recorded locales
- 108 count
- 2026-10-12 · Observed
- Languages
- 83 count
- 2026-10-12 · Observed
Methodology
Counts are inventory figures stated in Basis AI's 12 October 2026 dataset announcement. Hours, speakers, locales and languages are reported by the source; no re-sampling or independent quality audit is implied here.
Updated:
Subjects and places
Subject proposals
- Artificial intelligenceSuggested
- Subject suggested from the title or summary: AIhuggingface.co ↗