Skip to main content
Vertex AI Gemini Live runs Google’s speech-to-speech models through your own Google Cloud project, authenticated by a service account instead of an API key. This is a different connection from Gemini Live, which reaches the same family of models with an AI Studio key. Use this one when you need Google Cloud billing, IAM controls, or data kept in a specific region. It is also separate from Vertex AI Gemini, which serves Vertex Gemini as a standard language model.

​
What you need

​
Getting a service-account JSON

  1. In the Google Cloud Console, select or create a project.
  2. Enable the Vertex AI API (aiplatform.googleapis.com) on that project.
  3. Go to IAM & Admin → Service Accounts and create one.
  4. Grant it the Vertex AI User role (roles/aiplatform.user).
  5. Open its Keys tab, choose Add Key → Create new key, and select JSON.
  6. The file downloads once. Store it securely and treat it as a secret.
Upload the file as-is, including private_key and client_email. The project ID is read out of the JSON, so you do not type it — supplying a different one authenticates as the service account but targets another project, which fails mid-call with a 403.

​
Adding the connection

  1. Go to Connections and choose Vertex AI Gemini Live.
  2. Use Upload JSON to attach the service-account file.
  3. Save.
SentiVue Talk checks the credential with Google as you save. That confirms the service account authenticates — it does not confirm the project has the API enabled or the right role, which surface on the first call.

​
What you can select afterwards

Choose your model on Model Configurations. Model — Google’s native-audio Live model on Vertex. See Google’s Live API documentation. Voice — the same voices as the AI Studio Gemini Live connection, selected by name. Language — the same set as the AI Studio Gemini Live connection. Region — set per workflow, not on the connection, and it defaults to us-central1. The credentials and project ID stay on the connection, so they do not appear in the workflow’s model settings.

​
How a speech-to-speech call differs

A speech-to-speech model replaces the transcriber and the voice provider. When a workflow is set to realtime, those two settings no longer apply — the model hears the caller and speaks back itself.
  • The voice comes from the model. You cannot pair a Gemini Live model with a voice from another provider.
  • You still need a language model configured. It runs the work the realtime model does not — extracting variables and running post-call QA analysis. Without one, those features silently produce nothing. See Realtime.
  • Moving between nodes reconnects the session. Gemini Live cannot change its instructions mid-session, so a node transition restarts it. The reconnect waits until the agent has finished speaking, so the caller does not hear it cut off.
  • Context compaction does not apply. It is a text-mode feature and has no effect in realtime mode.
  • You still get a transcript. Both sides of the call are transcribed for the run record, so tracing and post-call analysis work as usual.

​
Troubleshooting

Paste the full service-account JSON key file, not an AI Studio API key. An API key belongs on the Gemini Live connection instead. Check that the file is the complete JSON object, including the private key.
Either the Vertex AI API is not enabled on the project, or the service account is missing the Vertex AI User role. A 403 that appears at call time but not on save usually means the role rather than the credential.
The workflow is not set to realtime. Switch it over on Model Configurations, then pick the Vertex AI Gemini Live model — the transcriber and voice settings stop applying at that point.