GPT-Live Voice YAML
Author GPT-Live Voice after selecting it with Voice Agent subtypes and applying common invariants. It is not a Prompt Voice setting and does not use a Hosted Voice Bridge.
Canonical Shape
services:
live-assistant:
host: azure.ai.agent
uses: [ai-project]
kind: voice
name: live-assistant
subProtocol: live
modelType: self_deployed
model:
id: gpt-live-1model.id must match the exact compatible deployment name.
Self-Deployed Model Golden Path
GPT-Live Voice has no managed-model branch. Apply this lifecycle to the GPT-Live model and to delegation.responses.model when Responses delegation is configured:
- Inspect the selected Foundry account for the exact compatible deployment name. Reuse an existing deployment without adding a duplicate entry.
- If it is missing, read Foundry Model Reference and resolve the exact version, format, supported SKU/capacity, lifecycle, and quota. State the billable/quota-consuming change, and obtain confirmation unless creation was explicit.
- Preserve existing project infrastructure and deployment entries, append the resolved deployment under
services.<ai-project-service>.deployments[], and setmodel.idordelegation.responses.modelto the exactdeployments[].name. - Run
azd provision --preview --no-prompt, confirm only the intended model infrastructure changes, then runazd provision --no-prompt. - Deploy the Voice service only after provisioning succeeds.
Do not use az cognitiveservices account deployment create in the normal skill flow; keep azure.yaml as the azd source of truth.
Optional Settings and Known Values
Preserve unrelated existing fields and omit unused optional fields rather than writing null.
instructions: Respond naturally and concisely.
store: false
audio:
format: { type: audio/pcm, rate: 24000 }
output:
voice: marin
greeting: { type: template, text: "Hello {{name}}" }
structuredInputs:
name:
description: Caller name
schema: { type: string }
required: false
defaultValue: Ada
policies:
- type: rai_policy
raiPolicyName: <rai-policy-resource-id>| Setting | Supported shape |
|---|---|
audio.format |
audio/pcm at 16000 or 24000; audio/pcmu or audio/pcma at 8000 |
audio.output.voice |
Nonblank built-in voice string, or { id: <custom-voice-id> } only |
instructions / store |
String / boolean |
greeting |
Exactly type: template plus nonblank text |
structuredInputs.<name> |
Optional description string, schema object, required boolean, and defaultValue |
policies |
At most one named rai_policy with full-resource-ID raiPolicyName; no invocationsModeration |
Delegation
delegation: { type: client }delegation:
type: responses
responses:
model: <responses-deployment-name>
instructions: Return concise backend guidance.delegation:
type: foundry_managed_agent
agentName: <existing-agent-name>
agentVersion: "<positive-version>"clientsupports onlytype.responsesrequiresresponses.modeland must omitagentName/agentVersion.foundry_managed_agentrequiresagentName, may include a positive numeric-stringagentVersion, and must omitresponses.
responses option |
Supported shape |
|---|---|
instructions |
String |
maxOutputTokens / parallelToolCalls |
Positive integer / boolean |
serviceTier |
auto, default, fast_tier_temp_pilot, flex, priority, or ultrafast |
reasoning |
effort: `none |
text.verbosity |
low, medium, or high |
toolChoice |
`auto |
tools |
{ type: web_search }, or function with name, optional description, parameters object, and optional boolean strict |
Constraints and Forbidden Fields
Forbidden or discriminator-incompatible fields are rejected even when empty or null. Do not author audio.input transcription/VAD/noise reduction/echo cancellation, audio.output.format/speed, conversationEngine, targetAgent, root voice, input/output schemas, output modalities, root tools/tool choice/output-token settings, avatar, handoff, include, or telephony.