Eleven v4 Turbo hits 150ms — voice agents got usable

The thing that made AI phone agents feel wrong was never the voice. It was the pause. On September 28 ElevenLabs shipped Eleven v4 and Eleven v4 Turbo, and v4 Turbo reports a median ~150ms time to first speech with ~100ms median inference latency. Human conversational turn-taking runs in roughly that range. For anyone who shelved a voice agent project because callers kept talking over dead air, the technical objection just got weaker.

What actually happened

Two models, one split. Eleven v4 is the expressive one, aimed at pre-recorded audio. Eleven v4 Turbo trades some of that for speed and is explicitly built for live agents. ElevenLabs publishes v4 Turbo at ~150ms median time to first speech and puts competing models between 262ms and 814ms on its own benchmark — treat a vendor's own comparison as a vendor's own comparison, but the absolute number is the part you can test yourself.

Language coverage moves from roughly 70 to over 90 languages, with the company calling out the largest quality gains in Japanese, Brazilian Portuguese, Mandarin and Cantonese. A single voice can speak any supported language while holding its speaker identity.

Instant voice cloning now needs 10 seconds of audio. Professional Voice Clones are supported for high-fidelity work. Inline audio tags carry over from v3 and got stricter adherence — [laughs], [sleepy drowsy voice], [said angrily in French accent], even environmental cues like [light rain] — and you can stack tags and have the model follow the sequence. IPA phoneme support improved, which matters more than it sounds if your catalog is full of product names no model has seen.

The models are live across ElevenAgents, ElevenCreative and the API. ElevenLabs did not publish v4 pricing in the announcement, and says 55% of its revenue now comes from enterprise customers, with an annualized run rate that moved from about $330M to over $600M inside the year.

Why faster voice AI matters for your business

Latency was the whole product. A 700ms gap before a reply reads as a bad connection or a bored employee. Around 150ms reads as a person. That single number decides whether callers stay on the line, and it is the first thing to measure in a pilot — end to end, on your network, with your telephony provider in the path, not from a vendor slide.

Turbo means you stop pretending the LLM is instant. v4 Turbo can start generating audio as soon as the language model behind it starts producing tokens. That is an architecture change, not a setting: your agent needs to stream, and any synchronous step you inserted — a CRM lookup, a stock check, a policy retrieval — now sits in the caller's ear as silence. Move those calls behind the first sentence or cache them.

10-second cloning is a consent problem before it is a feature. You can now clone a staff member's voice from a voicemail greeting. Get written permission, keep the recording and the authorization together, and decide up front whether your brand voice is a real employee or a synthetic one. One of those follows you when they quit.

90 languages is a market decision disguised as a model release. If a fifth of your inbound is Spanish or Cantonese and you handle it with a callback queue, the cost of covering it in-line just dropped. That is revenue, not tooling.

Disclose the bot. The model being good enough to pass is exactly why you say it is a bot at the top of the call. Regulation is tightening in that direction and customers punish the discovery harder than the disclosure.

Sources: Eleven v4, ElevenLabs, ElevenLabs' new v4 speech model supports more expression control and 90 languages, TechCrunch.