Prompt Voice YAML
Author a single managed Voice Agent after applying common invariants.
Managed Real-Time Model
modelType: managed uses the service-hosted model and does not require a project deployment:
services:
support-voice:
host: azure.ai.agent
uses: [ai-project]
kind: voice
name: support-voice
modelType: managed
model:
id: gpt-realtime
instructions: Respond naturally and concisely.Self-Deployed Model Deployment
For an azd-managed project, services.<ai-project-service>.deployments[] is the source of truth and azd provision creates or updates the model infrastructure. Prompt Voice references the exact deployment name through modelType: self_deployed and model.id.
Foundry infers native speech-to-speech versus cascaded STT -> LLM -> TTS from model.id; apply the same flow to realtime and text-model BYOM.
First inspect the resolved subscription, resource group, and Foundry account for the exact requested deployment name:
az cognitiveservices account deployment list --resource-group "<resource-group>" --name "<foundry-account>" --query "[?name=='<deployment-name>'].{name:name,model:properties.model.name,version:properties.model.version,format:properties.model.format,sku:sku.name,capacity:sku.capacity}" --output jsonVerify that the returned model is compatible with the requested Voice architecture. If it exists, preserve services.<ai-project-service>.deployments[], do not add a duplicate, and reference the exact name:
services:
advisor-voice:
host: azure.ai.agent
uses: [ai-project]
kind: voice
name: advisor-voice
modelType: self_deployed
model:
id: <existing-deployment-name>
instructions: Respond naturally and concisely.If the deployment is missing and the user intends creation:
- Read Foundry Model Reference and resolve the exact model version, format, supported SKU/capacity, lifecycle, and available quota. Never guess these values.
- State the billable/quota-consuming deployment plan and obtain confirmation unless creation was explicit.
- Preserve existing deployment entries and unrelated infrastructure, then append the resolved deployment to the existing
ai-projectservice. Setmodel.idto the exactdeployments[].name.
services:
ai-project:
host: azure.ai.project
deployments:
- name: <exact-deployment-name>
model:
format: <catalog-format>
name: <catalog-model-name>
version: "<catalog-model-version>"
sku:
name: <supported-sku>
capacity: <confirmed-capacity>
advisor-voice:
host: azure.ai.agent
uses: [ai-project]
kind: voice
name: advisor-voice
modelType: self_deployed
model:
id: <exact-deployment-name>
instructions: Respond naturally and concisely.Illustrative only: A verified realtime deployment used
gpt-realtimeversion2025-08-28,GlobalStandard, capacity1. Re-query the catalog, lifecycle, regional support, and quota before authoring; these values are not universal defaults.
For an existing azd project, edit deployments[] in place; do not rerun azd ai agent init. Do not use az cognitiveservices account deployment create, because it creates infrastructure outside the azd lifecycle. Run azd provision --preview --no-prompt, review that only intended model infrastructure changes, then run azd provision --no-prompt and deploy the Voice service.
Optional Configuration
When requested settings require YAML edits, load Voice Agent YAML settings. Preserve unrelated existing fields and edit only requested paths.
Common defaults include PCM at 24 kHz, Azure Speech transcription, server VAD, and audio output. Omit voice unless requested or required by an existing definition so azd applies its default.