ElevenLabs' Eleven v4 Turbo hits ~100ms, tops the voice arena

Real-time voice agents get cheaper and faster: inline audio tags, 10-second cloning, 90+ languages, and a REST/streaming API — with a two-week intro price.

Nowline SEP 30 6:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Turbo at ~100ms makes phone-speed agents real

    Eleven v4 Turbo posts a median inference latency around 100–150ms — versus 262ms for Cartesia Sonic 3.6 and 814ms for GPT-4o mini TTS — fast enough for natural, interruptible voice conversations. Intro pricing (first two weeks) runs $11 per million characters for Turbo and $22 for the full v4.

  • Audio tags + 10-second clones: a weekend build

    Write [laughs], [whispers], or [door slams] straight into your script, and spin up a usable voice from a 10-second sample — no fine-tuning. That's enough to prototype an IVR, a narrated app, or a game character with directable emotion and pacing in an afternoon.

  • Ranked #1 on Artificial Analysis' Voice Arena

    Eleven v4 took the top spot on Artificial Analysis' Provider Voice Arena (Elo 1,319), and roughly 75% of listeners preferred it in blind tests against five rival services. If you stayed on v3 or a competitor for quality, that reason just weakened.

  • 90+ languages, streaming API, telephony output

    v4 covers 90+ languages — Cantonese, Mongolian, and Odia included — with IPA pronunciation control, up to 10,000 characters per generation, and MP3/WAV/PCM/mu-law output for phone systems. It's live via REST and streaming plus ElevenAgents, and included at no extra cost on eligible Creator+ plans.