The First AI That Can Laugh

ElevenLabs has trained on a massive dataset of over 500,000 hours to present a new model that goes beyond simple voice synthesis to deeply understand emotion and context. This model grasps the emotion embedded in text to express joy, anger, sadness, and more, and in particular, when conveying the joy of victory, it naturally generates non-verbal sounds such as laughter.

Context comprehension has also been strengthened. By analyzing preceding and following sentences, the model maintains consistent emotional patterns even in long sentences, and it accurately distinguishes words like 'read' or 'minute' that are spelled the same but pronounced differently depending on context. It also reads abbreviations like FBI or symbols like $3tr naturally, in line with how they are actually spoken.

Key features and use cases:

• Emotion and context understanding: Grasps the meaning of text to provide rich expressiveness such as laughter and emphasis
• Spoken-language optimization: Natural pronunciation handling for abbreviations and symbols
• User feedback system: A feature is being developed that lets the model flag uncertain parts to users so they can directly correct them

This technology is expected to simultaneously reduce costs and improve quality across various fields, including turning news articles into audio, producing audiobooks with distinct character personalities, generating NPC voices in games, and advertising campaigns.