Eleven v4 brings finer voice direction, with Turbo for live conversations

01Direct the performance inside the script

ElevenLabs launched Eleven v4 and Eleven v4 Turbo on 28 September. This is a catch-up on that release, following its fuller product walkthrough on 30 September. The company says v4 improves emotional delivery, speaker consistency and the handling of inline audio directions.

Directions such as whispering, laughing or a change in mood sit alongside the dialogue. Audio tags existed before v4; the claimed improvement is how reliably and subtly the model follows them. The documentation still acknowledges imperfect tag adherence.

For a short product film, that means one script can distinguish a hesitant question from a confident reply. The useful creative unit is a performance with an intention, rather than several voices reading disconnected sentences. A game scene or a conversational advert could use the same approach. Those are suggested applications, not XANTRANS test results.

02Two models for different kinds of waiting

The model guide positions eleven_v4 for produced audio and multi-speaker dialogue, with a 10,000-character limit. It is available through the Text to Dialogue API. The eleven_v4_turbo variant is aimed at responsive voice experiences. Both support more than 90 languages.

ElevenLabs reports roughly 100 milliseconds median inference latency for Turbo, and about 150 milliseconds to first audible speech in its streaming comparison. The comparison removes network latency. Those figures are not a promise that an entire telephone agent will answer within 150 milliseconds.

A booking assistant must still understand the caller, choose a response and possibly wait for a calendar lookup. Faster speech generation can shorten one part of that chain. For an edited voiceover, a small wait is usually less important than whether the line lands correctly.

03Existing voices may sound different

Professional Voice Clones are supported. For an existing clone, the documented upgrade path is My Voices, then the plus control beside Eleven v4 to start fine-tuning. Clean source audio matters because the model can reproduce unwanted characteristics as well as the speaker's identity.

Cross-language accent handling also changes: by default, generated speech adopts the target language's native accent instead of carrying over the source language's accent. That can help localization, but it can change an intentionally accented character.

Stability and Similarity remain available controls. Style and Speed sliders are absent, and SSML is unsupported. A previous integration that depends on those controls therefore needs more than a model-name swap.

04Where to start

The launch announcement lists availability in ElevenCreative, ElevenAgents and ElevenAPI. For an existing project, compare the same short scene in the old and new model before regenerating a complete production.

Choose a passage containing the things your audience will notice: a name, a pause, an emotional turn and, where relevant, a language switch. Keep the version that serves the scene. A newer model can offer more control without making every existing voice performance obsolete.