ElevenLabs has announced two new speech models, Eleven V4 and Eleven V4 Turbo, aimed at more expressive speech and, in Turbo’s case, real-time use. The notable change is a split between high-fidelity long-form audio and lower-latency voice agents, with the source citing a median time to first speech of about 150 milliseconds for Turbo.
Watch Desk analysis
What happened
The release details come from the supplied Hacker News item, which says the models are available across ElevenLabs products and APIs. It describes V4 as suited to long-form work such as audiobooks, and Turbo as the option for conversational applications.
The listed changes include script-direction tags for shaping delivery, context stitching for longer scripts, more stable voices across regenerations and support for Professional Voice Clones. The item also says the models cover more than 90 languages and reports a top ranking for V4 on the Artificial Analysis Provider Voice TTS Arena leaderboard.
Why it matters
For creators, more control over delivery could mean fewer rounds of nudging a voice towards the right tone. For developers, Turbo’s reported latency makes live voice applications a more plausible fit, though the supplied account does not provide the benchmark details behind that figure.
Our read
This is a substantial update to a tool people use to make voices, not just a fresh model name on a menu. The useful test is whether direction tags and steadier regenerations make production less fiddly, while Turbo sounds natural at conversational speed. The leaderboard claim is worth noting, but a ranking is not a substitute for hearing the model in your own use case.
What to watch
- How the direction tags work across different voices and languages.
- Whether Turbo’s reported latency holds up in real applications.
- What access and pricing apply across the products and APIs.
Discussion spark: For voice tools, which matters more in practice: fine control over emotional delivery, or low enough latency for a conversation to feel live?
Sources and evidence
- ElevenLabs releases v4 speech model () (28 September 2026, 18:49 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.