Gemini Live & Live Extended Thinking: Sub-200ms Conversational Voice APIs

Overview #

Gemini Live and Gemini Live Extended Thinking provide real-time, bidirectional conversational speech-to-speech APIs over WebRTC, allowing AI voice agents to reason deeply while conversing naturally with acoustic turn-taking.

Native Speech-to-Speech without Text Bottlenecks #

By bypassing text transcription cascades (VAD -> ASR -> LLM -> TTS), Gemini Live communicates natively in audio tokens over WebRTC media channels. Total latency drops to 180-250ms, delivering human conversational cadence with pitch variation, laughter, and natural pacing.

Live Extended Thinking in Conversational Audio #

Gemini 3.8 Live Extended Thinking allows the model to perform internal reasoning tokens during a live audio stream without generating awkward dead silence. The model produces subtle conversational fillers or backchannel confirmations ('Let me calculate that for you...') while its internal reasoning graph resolves the answer.

Acoustic Barge-In & Interruption Handling #

When a user speaks while Gemini Live is responding, the server immediately truncates the outgoing audio stream and discards downstream buffers, enabling true conversational interruption.

Code Example: Establishing a Gemini Live WebRTC Audio Session #

async function startGeminiLiveSession(token: string) {
  const pc = new RTCPeerConnection();
  const stream = await navigator.mediaDevices.getUserMedia({ audio: true });
  stream.getTracks().forEach((track) => pc.addTrack(track, stream));

  const audioEl = document.createElement("audio");
  audioEl.autoplay = true;
  pc.ontrack = (e) => { audioEl.srcObject = e.streams[0]; };

  const offer = await pc.createOffer();
  await pc.setLocalDescription(offer);

  const res = await fetch("https://api.apihundred.com/v1/realtime?model=gemini-3.8-live", {
    method: "POST",
    headers: { Authorization: `Bearer ${token}`, "Content-Type": "application/sdp" },
    body: offer.sdp,
  });

  const answerSdp = await res.text();
  await pc.setRemoteDescription({ type: "answer", sdp: answerSdp });
  console.log("Gemini 3.8 Live Voice Session Active");
}

Frequently Asked Questions #

Q: What is Gemini Live Extended Thinking?

It is a dedicated real-time voice model that reasons step-by-step during live audio streaming, allowing voice agents to solve complex logic without losing conversational cadence.

Q: Does Gemini Live require WebSockets or WebRTC?

Gemini Live supports WebRTC for lowest latency (UDP transport) as well as bidirectional WebSockets for server environments.

Q: Can Gemini Live execute tools during voice conversations?

Yes, it supports real-time function calling, allowing the voice assistant to query databases or invoke APIs mid-dialogue.