Providers¶
Two speech-to-speech providers ship today. They are genuinely different, and the honest version is that quality and responsiveness pull in opposite directions — there is no setting that gives you both.
OpenAI Realtime¶
The most conversational of the two. Interruption handling is excellent: cut in mid-sentence and the agent stops immediately and picks up where you took it. Turn detection is quick and rarely talks over you.
The weakness is non-English audio. Russian in particular is intelligible but audibly accented, in a way callers notice.
Gemini Live¶
Markedly more natural non-English speech, across 70+ languages. If your callers speak Russian, this is usually the one to try first.
The trade is interruption. Barge-in is handled entirely on Google's side and reported after the fact, so cutting the agent off is less crisp than with OpenAI — expect it to finish a word or two more than you would like.
Choosing¶
| If you care most about | Choose |
|---|---|
| English conversations that feel natural to interrupt | OpenAI Realtime |
| How the agent sounds in Russian or another non-English language | Gemini Live |
Both are configured the same way and cost is charged the same way — by call duration, at the per-minute price shown next to the model. Switching is a dropdown; nothing else about the agent changes.
Custom voices¶
OpenAI lets you use a voice of your own. You create it once against OpenAI's API —
upload a consent recording from the voice actor, then a sample of up to 30 seconds — and
get back an id like voice_1234. Paste that id into the Voice field instead of
picking a built-in name, and calls will answer in that voice. The feature is limited to
accounts OpenAI has enabled for it (their sales team gates access), and an organization
may hold at most 20 voices.
Gemini Live has prebuilt voices only. Google's voice cloning — Chirp 3: Instant Custom Voice — is part of Cloud Text-to-Speech, is itself allow-listed, and does not reach the Live API, so there is nothing to paste here.
This is why the voice field behaves differently between the two: for OpenAI you can type as well as choose, for Gemini you choose from the list.
Your own API key¶
By default a call uses the Assemblix system key. Select a credential in the Voice section to bill the provider directly instead. The credential type has to match the provider — an OpenAI key for OpenAI, a Gemini key for Gemini.