VOICE AGENT INTEGRATION
Gemini Live is Google's speech-to-speech API, running speech recognition, the model and the voice inside one session, and it connects to Vobiz for telephony: Vobiz provides the phone numbers and carries the calls, bridged to Gemini by an open-source reference service that Vobiz publishes. The audio formats line up exactly at both ends, so nothing is resampled between the caller and the model.
Before you start
This one is code you run rather than a provider you select. Vobiz publishes an open-source bridge in Python that joins a Vobiz media stream to a Gemini Live session, and you deploy it somewhere Vobiz can reach over HTTPS and WSS — a tunnel is fine while you are building. It handles inbound and outbound without caring which direction the call came from. You will need a Gemini API key and a Vobiz number.
Why this pairing
Gemini Live holds a spoken conversation in one session: it hears the caller, decides what to say and says it, with its own sense of when someone has stopped talking. There is no transcription step to tune and no separate voice vendor to pay. What it cannot do is reach a phone.
The usual cost of bridging a model like that to a phone line is conversion — the caller's audio resampled on the way in, the model's speech resampled on the way out, each pass costing a little quality and a little time. Gemini is fixed at one rate in and another out. Vobiz sets the two directions of a stream independently. Line them up and the conversions disappear.
What you get
The bridge is published. What is left is a key, a number and a prompt.
Speech recognition, the model and the voice all run inside a single Gemini session on a single API key. No separate transcription vendor, no separate speech vendor, and no turn-taking logic of your own to write and then debug on a live call.
The two directions of a Vobiz stream are configured independently, which means they can be matched to what the model actually wants — one rate arriving, a different rate going back. Set both correctly and the caller's audio reaches the model untouched and the model's speech reaches the call untouched. Other rates work and the bridge converts them; this is the path where it does not have to.
When a caller talks over the agent, Gemini reports the interruption and the bridge tells Vobiz to discard the audio still queued on the line. Vobiz confirms the flush. Queued speech can run to seconds, so without both halves the caller keeps hearing a reply the model has already abandoned.
Sending audio is not the same as the caller hearing it. After each turn the bridge asks Vobiz to confirm playback, which is what lets the agent finish its closing line before the call ends rather than being cut off mid-word — and makes a reply nobody heard visible in the log rather than only in the call.
Local, mobile and toll-free inventory across 130+ countries, including regulated formats where a market demands them — India's 140-series and 1600-series among them. Point a number's answer URL at the bridge for inbound, or place an outbound call with the same URL; the bridge does not distinguish.
The repository ships a mock client that opens the socket the way Vobiz does, speaks a prompt, and writes the agent's reply to a file — so the bridge, the model and the formats can all be verified before a number is ever dialled, and in continuous integration afterwards.
How it works
Pick a model, fill in a file, run it, dial.
Model names on the Live API change often, and choosing one your key cannot reach fails at connect time with a bare socket close and no useful message. The repository includes a script that lists exactly which ones are available to you.
Clone it, create a virtual environment, install the requirements and fill in the environment file: your Gemini key, the model, the voice, the greeting and system prompt, and your Vobiz credentials. The audio settings already carry the matched defaults.
A self-test runs dozens of checks without placing a call. Then start the bridge; it prints the answer URL, the socket address and the formats it will use in each direction.
Dial out with the included script, or point a Vobiz number's answer URL at the bridge for inbound. Transcripts and stream events are readable from the bridge while the call is still running.
The full guide covers the format matrix, the stream protocol, barge-in and checkpoints, the verified call measurements and a behaviour reference mapping each symptom to its cause.
Read the integration guide
Get started
Create a Vobiz account and get a number, clone the bridge, add your Gemini key.
The self-test tells you it works before you dial anything.
FAQ
CRM and helpdesk integrations: Salesforce · ServiceNow · ClickUp · Pipedrive · Expedify · HubSpot · Freshdesk · Zendesk
Voice AI platforms on Vobiz: Bolna · Vapi · Retell AI · LiveKit · ElevenLabs · See all integrations
Connecting by API instead of the dashboard: Retell AI API guide · ElevenLabs API guide
Gemini Live is Google's API for real-time spoken conversation with a Gemini model. Learn more in Google's Live API documentation. Vobiz is not affiliated with Google; this integration uses the public Live API. Model names and defaults reflect the reference bridge as of 21 October 2026.
We use cookies to run this site and, with your permission, to understand how it's used. Read our Cookie Policy.
