The Pulse
Basis Releases 525 Hours of Expressive Speech Across 108 Locales
Basis has released Global Emotions, a 525.17-hour speech dataset spanning 108 locales, 83 languages and 10,420 speakers. The audio is licensed for commercial and research use, but access requires a reviewed request and human-validated trans
AI.info Team · Written with AI ·
Basis calls its new Global Emotions speech dataset free for commercial and research use, but access comes with a gate: prospective users must request it, and a person reviews each request. The company also charges separately for human-validated transcripts, which are not included with the audio. Released October 12, the collection pairs a broad license with limits on how users obtain and handle recordings.
One speaker, one line, several deliveries
Global Emotions contains 525.17 hours of audio across 137,696 recordings, grouped into 54,575 matched series from 10,420 speakers. Its 108 locales cover 83 languages and 90 countries. Audio files are mono WAV at 48 kHz and 16-bit.
Each series keeps the speaker and scripted line constant while changing the performance. Emotion series include neutral, medium and full-intensity readings; delivery series pair a neutral reading with a full delivery, such as whispering or singing. “The same voice actor reads the same line according to three different emotional levels (e.g. neutral, angry, & very angry),” writes Alexis Sursock in the Basis Global Emotions announcement.
The paired format lets model builders compare how emotion or delivery changes a voice without also changing the speaker or words. The collection includes 12,141 recordings with a scripted non-verbal cue, such as a sigh, scoff, gasp or laugh; the dataset card lists 13 cue types. Basis says the cues appear in the instructed scripts, not in human-verified transcripts.
Broad license, reviewed access
Basis says contributors recorded through its consumer applications, received payment and consented to licensing their recordings for AI training. The dataset uses stable, salted speaker identifiers and excludes names, account identifiers and contact details; demographic information is self-reported. The license permits commercial and research use, including training and selling models, but bars redistribution of the recordings themselves and attempts to identify, contact or impersonate speakers.
Users must request dataset access, and Basis says a person reviews requests. The release is listed on Hugging Face at 168 GB. Basis’s open-data page presents the collection as a newly released dataset, while the dataset card supplies the detailed access and license terms.
Scripts are not verified transcripts
The text field in each record, called given_line, contains the script a speaker was asked to read; Basis warns that speakers may depart from it. Human-validated transcripts are not part of the open release. The announcement lists them separately at $22.7 per audio hour.
Basis says it selected recordings using ratings from native speakers alongside internal models for following the requested emotion or delivery and for microphone quality. Selection was balanced across delivery instructions and genders, with a cap of 35 minutes per speaker and ten series per line. The audio is supplied as recorded, without later denoising, enhancement, loudness normalization or resampling.
Regional varieties are part of the design
The 108 locales distinguish regional varieties within languages, rather than treating each language as a single voice category. The release includes Brazilian and European Portuguese and Kenyan and Tanzanian Swahili, as well as multiple varieties of Arabic, English and Spanish. Basis’s dataset card reports the hours, recordings, series and speaker counts for each locale.
That detail gives researchers a way to test whether an expressive speech model handles both delivery and regional variation, while the matched readings offer controlled comparisons within a speaker. But users working with text labels should account for the gap between the prompted script and what was actually spoken: the open dataset supplies the former, not verified transcripts of the latter.