Made by AI, Directed by David Stein

Search
AI & Creative Tools

Google's New Live Voice Models Make Talking Installations Cheap Enough to Actually Build

Gemini 3.8 Live is voice-to-voice, sees through a camera, calls tools mid-sentence and switches between 97 languages — at half a cent a minute to listen and under two cents to speak.

3 min read

Anyone who has built an installation that talks to visitors knows the architecture, and knows why it usually disappoints. You run speech-to-text, send the text to a language model, wait, then run text-to-speech. Each hop adds delay, and the total lands somewhere around two or three seconds. People read that pause as the machine not understanding them, and they walk away.

Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, released September 15, are Google’s latest attempt to remove those hops. They’re voice-to-voice models: audio goes in, audio comes out, with no round trip through text in the middle.

What they do

  • Real-time spoken conversation, with interruption handling rather than turn-taking
  • Visual grounding: the model understands what a connected camera sees while you talk to it
  • Tool and API calls mid-conversation, without going silent while they run
  • 97 languages, switchable mid-sentence
  • Two tiers: 3.8 Live for scale and cost, 3.8 Live Extended Thinking for tasks that need multi-step reasoning

Availability is the Gemini API and Google AI Studio for developers, with enterprise access in private preview, plus rollout across Search Live, Gemini Live and Workspace.

The pricing is the story

$0.005 per minute of audio input. $0.018 per minute of audio output.

Work through what that means for an installation. A gallery piece that talks for six hours a day, with the model actually speaking maybe a third of that time, costs roughly $0.40 a day in output plus a couple of cents to listen. Call it $12 a month for a piece that holds conversations with everyone who walks up to it.

Conversational installations have historically failed on two things: latency that made them feel broken, and running costs that made a long exhibition run untenable. A number that low reframes the question from “can we afford to leave it on” to “is the piece any good.”

Where it fits, and where it doesn’t

The camera grounding is the most interesting part for artists. A voice model that can see means a piece can respond to what a visitor is holding, wearing or pointing at, without you building a separate vision pipeline. Tool calls during speech mean it can query a database, control lights or trigger DMX while it’s still talking, instead of freezing mid-sentence.

The tradeoffs are the usual ones, and they are real:

  • It’s a cloud API. Your installation needs reliable internet, and audio from your visitors goes to Google. For a gallery that’s a consent-notice problem; in some contexts it’s a non-starter.
  • You don’t control the voice or the model’s manner. Work that needs a specific character will fight the defaults.
  • Prices and models change. Anything built on a hosted API inherits that risk, which is why the local-first alternatives we keep covering matter even when they’re worse on paper.

For a piece that must run offline or keep audio private, a local stack (Whisper plus a small LLM plus Kokoro or Piper) is still the answer, at considerably more integration work and worse conversational feel. For everything else, this is now cheap enough to prototype this weekend.

Related Reading