Gemini Live on a real phone call with the Wixzel Voice API

What do you need before you start?

• A Wixzel Voice API key with credit. The minimum top-up is $5.
• A SIP trunk at your carrier and a number on it that the carrier lets you present as caller ID.
• curl, jq and uuidgen.

How do you build this on Wixzel Voice, step by step?

1. 01Connect your SIP trunk

   Set your key and the API base once, then create the trunk. tee keeps the full response, because you need two values from it besides the id:

   shellCopy

   export WIXZEL_API_KEY="wv_live_..."
   API=https://api.voice.wixzel.com/v1
   
   TRUNK_ID=$(curl -s $API/sip-trunks \
     -H "Authorization: Bearer $WIXZEL_API_KEY" \
     -H "Content-Type: application/json" \
     -d '{
       "name": "My carrier",
       "host": "sip.your-carrier.example",
       "username": "your-sip-username",
       "password": "your-sip-password"
     }' | tee trunk.json | jq -r .id)
   
   jq -r '.platform_ip, .origination_uri' trunk.json

   Leave out username and password if your carrier authorises by IP. Either way, allowlist platform_ip at the carrier so it accepts calls from the platform; origination_uri is where the carrier should send inbound calls, which this guide does not need.

2. 02Register the number you will call from

   Use your own number, in E.164 form:

   shellCopy

   NUMBER_ID=$(jq -n --arg trunk "$TRUNK_ID" \
     '{phone_number: "+14155550100", sip_trunk_id: $trunk}' |
     curl -s $API/phone-numbers \
       -H "Authorization: Bearer $WIXZEL_API_KEY" \
       -H "Content-Type: application/json" \
       -d @- | jq -r .id)

3. 03Create the Gemini Live agent

   The voice is one realtime model rather than three stages:

   shellCopy

   AGENT_ID=$(jq -n --arg number "$NUMBER_ID" '{
       name: "Gemini Live",
       system_prompt: "You are a friendly assistant on a phone call. Keep answers to one or two sentences.",
       opening_message: "Hi, this is an automated assistant. Do you have a minute?",
       outbound_phone_number_id: $number,
       voice: { realtime: { model: "google/gemini-live", voice: "Charon" } }
     }' |
     curl -s $API/agents \
       -H "Authorization: Bearer $WIXZEL_API_KEY" \
       -H "Content-Type: application/json" \
       -d @- | jq -r .id)

   voice.realtime hands the whole turn, listening, reasoning and speaking, to Gemini Live, so there is no stt, llm or tts block; the API refuses a voice that names both shapes. Charon is one of Gemini's prebuilt voices, and GET /v1/engines/gemini_live/voices lists them all with Google's one-word descriptions.

   outbound_phone_number_id gives the agent a caller ID, which is what lets the next request leave out from_number_id.

4. 04Place the call

   Replace the to number with the phone you want to ring:

   shellCopy

   CALL_ID=$(jq -n --arg agent "$AGENT_ID" '{to: "+14155551234", agent_id: $agent}' |
     curl -s $API/calls \
       -H "Authorization: Bearer $WIXZEL_API_KEY" \
       -H "Idempotency-Key: $(uuidgen)" \
       -H "Content-Type: application/json" \
       -d @- | jq -r .id)

   The API answers 201 with status: "queued" at once, and the phone rings a moment later. Idempotency-Key is required here: a network timeout tells you nothing about whether the call was placed, and a retry with the same key returns the original call instead of dialling twice, so keep the key if you script retries. A call that cannot fund its first 15 seconds is refused with 402 before anything is dialled.

5. 05Read what happened

   The fifth request reads the call back:

   shellCopy

   curl -s $API/calls/$CALL_ID \
     -H "Authorization: Bearer $WIXZEL_API_KEY" | jq

   Example response, shortenedCopy

   {
     "id": "6a96a3ead6e886d42462dd3e",
     "object": "call",
     "session_id": "sip-5b0c7c0e-9d1f-4c52-8a9e-2f0d3a4b6c7d",
     "status": "completed",
     "direction": "outbound",
     "channel": "phone",
     "engine": "gemini_live",
     "to": "+14155551234",
     "duration_seconds": 94,
     "cost_micros": 73007,
     "failure_code": null,
     "failure_reason": null,
     "transcript": [
       { "role": "assistant", "content": "Hi, this is an automated assistant. Do you have a minute?" },
       { "role": "user", "content": "Sure, go ahead." }
     ]
   }

   status moves from queued through ringing and in-progress to completed, or ends as failed, busy or no-answer. cost_micros is written when the call settles (1,000,000 is $1.00) and stays null on a call that never connected, because such a call is not charged. When a call fails, failure_code is the carrier's Q.850 cause (17 busy, 19 no answer, 21 rejected by the carrier) and failure_reason says what to do about it. GET /v1/usage/events?session_id=… itemises every charge.

What does this build cost on Wixzel Voice?

Which Wixzel Voice limits apply to this build?

• Phone calls need your own SIP trunk. Wixzel Voice does not sell, port or hold numbers.
• A test call from POST /v1/agents/{id}/test-call speaks its one phrase through the classic voice pipeline, because Gemini Live has no single-phrase mode. To hear the Gemini voice itself, place an ordinary call.
• Five calls can be live at once per account by default, raised on request.
• No latency figure is published: a realtime engine answers as fast as the model itself does. Wixzel Voice is in alpha, and behavioural changes ship behind a dated Wixzel-Version header.