On October 12, Basis introduced Global Emotions: 525 hours of custom-recorded speech from 10,420 people across 108 locales covering 83 languages. Its distinctive feature is pairing: the same speaker reads the same line at different emotional intensities or delivery styles. This supports studying prosody while holding words and speaker identity constant.
Basis says contributors are paid, consent to recording and licensing for AI training, and that native speakers rate recordings. Those are supplier statements, not an independent audit of every consent. Speaker identifiers support speaker-disjoint training and test sets; a style label must be read together with intensity because a series’ neutral reading also carries the series label.
Hugging Face access is restricted and requires approval. Basis advertises commercial and research use under its licence, but the licence text itself was unavailable anonymously during the check, so no additional freedoms are assumed. Another important limit is that validated transcripts are not included in this release: given_line is the requested script, not a guarantee of what was actually spoken. This is directed performance data, not a measurement of spontaneous emotions in real-life situations.