ElevenLabs has officially launched Eleven v3, expanding support to over 70 languages and introducing inline audio tags for precise emotional control, alongside a new Text-to-Dialogue feature for seamless multi-speaker generation.
In August 2026, ElevenLabs officially rolled out Eleven v3, dubbing it their most expressive AI voice model to date. Building on the capabilities of its predecessors, the v3 model broadens its language support from 29 to over 70 languages, making it a highly versatile tool for global content creators, developers, and enterprises.
A standout addition to Eleven v3 is the introduction of inline audio tags. Previously limited to basic tags like pauses, users can now embed descriptive tags directly into the text—such as [laughs], [whispers], [sighs], and emotional cues like [sad] or [excited]. This allows for a much broader dynamic range and precise control over the direction, delivery, and emotional depth of the generated speech, reducing the need for trial-and-error prompting.
Another major feature introduced alongside v3 is Text-to-Dialogue. This capability weaves multiple voices together in a single generation pass, creating seamless interactions between speakers. The model intelligently matches prosody and emotional range across different voices, handling interruptions and shifting moods with limited prompting, making it ideal for generating engaging podcast segments, audiobooks, and conversational agents.
While Eleven v3 sets a new standard for narrative and high-fidelity character work, the company acknowledges it is not optimized for real-time, ultra-low-latency applications due to its larger architecture. For those use cases, ElevenLabs continues to recommend its Flash v2.5 models. Meanwhile, Eleven v3 is accessible via their web app and public API, cementing its position as a premium choice for studio-quality synthetic audio.
