Why this matters
Emotion and delivery control are central to making speech models sound natural and to improving paralinguistic understanding. This release supplies a large, paired, multilingual collection designed so models can learn how the same scripted line is spoken at different emotional intensities and delivery styles across many languages and regions.
What Sets It Apart
- Paired series design: the same speaker reads the same line at neutral and one or two instructed intensities, enabling controlled comparisons for fine-grained emotion modeling and contrastive training.
- Broad multilingual coverage: spans 108 locales (83 languages, 90 countries), including many underrepresented languages, which helps evaluate cross-lingual generalization of expressive speech and emotion classifiers.
- Scale and metadata: ~525 hours, 137,696 recordings, 10,420 speakers with locale/gender/age_band/steer/intensity metadata and scripted non-verbal cues, useful for supervised TTS, voice conversion, and emotion recognition research.
- Permissive training license: contributors consented to model training and voice synthesis use, allowing commercial and research model development while forbidding redistribution of raw audio and speaker identification/impersonation.
Who It's For & Tradeoffs
Great fit if you are training or evaluating expressive TTS, voice-cloning (with consent), speech emotion recognition, or cross-lingual paralinguistic models and need controlled paired examples across many languages. Look elsewhere if you need spontaneous conversational speech, long-form recordings, or raw audio redistribution rights—this dataset uses scripted lines (not natural dialogue) and the license disallows reselling or redistributing the recordings themselves. Also note demographic labels are self-reported and scripted format limits modeling of naturally occurring emotional dynamics.
