ElevenLabs — clip voices

eleven_multilingual_v2 · 21 voices · 8 kHz · 5,901 of 10,000 monthly characters

The recording-consent notice — the first thing every caller hears, and the hard case for any speech engine: fifteen seconds of uninterrupted speech with no breath and no self-correction to hide behind. Rendered through the exact path that ships, so what you pick is what deploys.

What's live right now — OpenAI marin

Shortlist — American, front-desk plausible

My pick of the account's voices: American accent, professional or conversational register. Accent is not taste here — the clips tell callers to dial 9-1-1, so the practice is American, and a British receptionist is the first thing an interviewer would notice.

Everything else on the account

The rest of the free catalogue, included so the choice is yours rather than mine. These skew toward narration, social-media energy, and character work — several are British, and a couple run over 19 seconds on a clip you want short.

How to pick

Why not the voice from the library

The library voice you linked returns HTTP 402: “Free users cannot use library voices via the API. Please upgrade your subscription to use this voice.” That is a hard paywall on the API path specifically — the web app will happily preview it, which is why it looks available. Everything on this page is a premade voice, which the free tier does allow through the API.

One thing this does not fix

The agent still speaks through OpenAI Realtime as marin, so picking an ElevenLabs voice here puts two voices back on the call — the thing you asked me to remove earlier. The clean fix is routing the agent's text through ElevenLabs too, which keeps the guard layer intact and gives one voice everywhere. Worth deciding after you hear these, not before.