Implement text analysis solutions 4.2 Implement speech solutions
You have a text-based prompt agent in Foundry Agent Service.
Study guide skill: Integrate speech as an agent modality, including custom speech models
Question B3-025
You have a text-based prompt agent in Foundry Agent Service. You want a managed real-time voice experience in which the model, instructions, voice, greeting, and turn-taking are managed together and saved as agent versions, without hosting any voice orchestration code. What should you do? (Voice-based agents are currently in preview.)
Select one answer.
Answer: D
A voice-based prompt agent is a separate agent whose definition holds the model, instructions, voice, greeting, and turn-taking, saved and tested as agent versions, while Foundry manages voice orchestration through Voice Live. You can't change an existing text agent's interaction mode to voice, so you create a new agent alongside it. With Voice Live connected to a text agent, you tune speech in Voice Live and manage the agent's instructions and tools separately, and hosting your own orchestration contradicts the requirement.
Deep dive
How it works
Foundry now has two ways to add voice to an agent, and the migration page compares them side by side:
- Voice Live with a Foundry text agent: Voice Live converts speech to text for the agent and synthesizes its response. You tune speech through Voice Live and manage the text agent's instructions and tools separately.
- A Foundry voice agent (a voice-based prompt agent, in preview): a separate agent with kind: voice. It still uses Voice Live for the voice runtime, but you manage the model, instructions, voice, greeting and turn-taking together in Foundry, and save and test changes as agent versions.
The requirement lists exactly the second column: model, instructions, voice, greeting and turn-taking managed together and versioned, with no orchestration code to host. The Agent Service overview says the same: create a voice-based prompt agent, configure the voice experience, and connect through Voice Live without hosting the orchestration code yourself.
How you get there matters. Migration creates a new agent alongside the original, because you can't change an existing text agent's interaction mode to voice. You can keep the text agent and even reuse it as a specialist subagent of the voice agent.
The other options, and when you'd use them
- Change the text agent's interaction mode to Voice: Not possible. You can't change an existing text agent's interaction mode to voice; you create a new agent alongside it.
- Connect the text agent to Voice Live and save voice settings as agent versions: With that integration, you tune speech in Voice Live and manage the text agent's instructions and tools separately. Voice, greeting and turn-taking aren't saved together as agent versions.
- Deploy gpt-realtime and host your own WebSocket service: That's the custom path where your code owns orchestration. The requirement says no voice orchestration code to host.
Exam tip
"Voice settings versioned with the agent, no orchestration code" = a new voice-based prompt agent. Existing text agents can't be switched to voice.
References
Last reviewed: