ElevenLabs Says Its New Voice Model Responds in ~150ms. The 'Acting' Claim Is the Harder One to Verify.

ElevenLabs, a text-to-speech vendor whose synthetic voices power audiobooks, dubbing and customer-service agents, released Eleven v4 and Eleven v4 Turbo on September 28, 2026, the company said. The Turbo variant's median time-to-first-speech lands at roughly 150 milliseconds in ElevenLabs' own testing, per the announcement's latency footnote.

That latency figure is the part an outside benchmark can corroborate. Artificial Analysis's Provider Voice leaderboard currently places Eleven v4 in a latency tier no other commercial voice model occupies, with an Elo of 1316±18 across 1,731 samples. The leaderboard measures how fast a model begins speaking, not how it "acts" — and that distinction is doing real work in ElevenLabs' marketing.

ElevenLabs is pitching the release as a generational shift from models that "read" text to ones that "interpret" it before synthesis, with claimed gains in contextual delivery, audio-tag following and speaker consistency across more than 90 languages. The framing is the vendor's, not an independent technical finding, and the same benchmark does not measure expressiveness or interpretation. ElevenLabs' own blind paired tests report roughly 75% listener preference against named competitors, though the announcement does not disclose sample size.

What listeners and developers can plan against today is the latency tier. Whether the interpretive layer earns the same weight is the question the release invites but does not yet answer.