Eleven v3 released

ElevenLabs has officially released Eleven v3, which offers richer expressiveness. This model goes beyond simple audio quality improvements to focus on solving the 'expressiveness' problem that was a limitation of previous models, including emotional escalation, interrupting during conversation, and natural interaction.

Key updates include the following:

• Audio tags: Tags such as [excited], [whispers], and [sighs] can be used to finely control tone, emotion, and non-verbal reactions.
• Dialogue mode: Supports multiple speakers conversing with natural breathing and interruptions.
• Support for 70+ languages: Supports a variety of languages with high global demand.
• Improved text comprehension: More accurately captures the emphasis and rhythm of input text.

Using the newly introduced Text to Dialogue API, you can structure speaker-by-speaker dialogue order through an array of JSON objects, and the model automatically handles speaker transitions and emotional changes.

However, Eleven v3 requires more sophisticated prompt engineering than previous models, and due to high latency, it is optimized for media production environments such as video production or audiobooks rather than real-time conversation use. For real-time services, using the v2.5 Turbo or Flash models is recommended.