Using Azure OpenAI GPT Realtime

How Azure Differs from Standard OpenAI

When you use a standard OpenAI account, your VNS CallStart response provides an openAIToken and VNS connects to OpenAI's shared API endpoint. With Microsoft Azure OpenAI, you instead provide a full aiUrl object that points to your own Azure-hosted resource.

Prerequisites

Before configuring VNS, you need the following set up in Microsoft Azure:

1. 1 An Azure subscription with access to Azure AI Foundry or Azure OpenAI Service.
2. 2 An Azure AI Foundry project or Azure OpenAI resource created in a region that supports GPT Realtime models. Visit Microsoft AI Foundry to create one.
3. 3 A deployed GPT Realtime model — go to Models + endpoints → Deploy base model in the Foundry portal and deploy a gpt-realtime variant.
4. 4 Your resource's WebSocket endpoint URL and API key — both available from the Azure portal (see below).

Finding Your Endpoint and API Key

In the Microsoft AI Foundry portal, navigate to your project and open Models + endpoints. Select your deployed GPT Realtime model. The endpoint details panel shows the base URL and your API key.

VNS needs the WebSocket (wss://) form of the endpoint with the model appended as a query parameter. The format is:

wss://<your-resource>.services.ai.azure.com/openai/v1/realtime?model=<model-name>

For example, a resource named monroe-7384-resource using model gpt-realtime-1.5 would be:

wss://monroe-7384-resource.services.ai.azure.com/openai/v1/realtime?model=gpt-realtime-1.5

Getting Your API Key

From the same endpoint details panel, copy the Key value (a long alphanumeric string). This is passed in the api-key header — not as a Bearer token.

CallStart Webhook Configuration

Use the aiUrl object in your CallStart response instead of openAIToken. The aiUrl.url is your Azure WebSocket endpoint and aiUrl.headers carries the api-key.

{
  "VoiceElements_Instructions": {
    "aiUrl": {
      "url": "wss://<your-resource>.services.ai.azure.com/openai/v1/realtime?model=gpt-realtime-1.5",
      "headers": {
        "api-key": "<<vns-api-token:YourAzureKeyName>>"
      }
    },
    "VoiceElements_Instructions": "You are a helpful AI voice assistant...",
    "maxSeconds": 300,
    "customerCallId": "{{sessionId}}"
  }
}

The <<vns-api-token:YourAzureKeyName>> syntax is a VNS placeholder — replace YourAzureKeyName with the profile name you gave the key when you stored it in the VNS portal. If you want to test quickly with a raw key, you can paste the key value directly — just remember to move it to a placeholder before going to production.

With a Raw Key (for Initial Testing)

{
  "VoiceElements_Instructions": {
    "aiUrl": {
      "url": "wss://monroe-7384-resource.services.ai.azure.com/openai/v1/realtime?model=gpt-realtime-1.5",
      "headers": {
        "api-key": "a1b2c3d4e5f6a1b2c3d4e5f6a1b2c3d4"
      }
    },
    "VoiceElements_Instructions": "You are a helpful AI voice assistant.",
    "maxSeconds": 300
  }
}

Supported Models

VNS works with any GPT Realtime model that Microsoft Azure makes available. The model name goes in the ?model= query parameter of your endpoint URL. The following models are currently available for global deployments in Azure AI Foundry:

Model availability varies by Azure region. Check the Microsoft Azure documentation for the current list of supported regions and model versions.

Troubleshooting