GPT Live voice agents that take action
Assistable now supports GPT Live voices for AI assistants, with 22 voices to audition and a separate model for chat and tool actions. Your team can choose how the agent sounds while keeping business lookups connected to the tools you configure.
Feature released
“Do you serve my area?”
A spoken request. A useful answer.A caller asks whether your team serves their area. A useful voice agent can ask for the right detail, check your business information, and explain the answer aloud. The voice matters. So does what happens behind the answer.
The result is a voice experience built around both conversation and the work a customer needs done.
Give your agent a voice that fits the conversation
Open the assistant’s voice drawer and explore the OpenAI GPT Live voices. Listen to the auditions before choosing the voice that suits your brand and the conversations your team handles.
A service desk may want a measured, reassuring delivery. A sales assistant may suit a more energetic voice. Listen for clarity, tone, and how comfortably you can follow the spoken sample. With 22 options, your team can compare voices before committing to one.
An audition helps you choose the sound. It plays a fixed sample; it does not run your assistant’s instructions, look up business data, or exercise its connected tools. Once you have a shortlist, test the selected voice in a published assistant call with your actual workflow.
Connect spoken questions to useful business answers
GPT Live handles speech and turn-taking. When a request needs a lookup or another tool action, the assistant delegates that work to its configured reasoning model and tools, then speaks the result.
That gives teams two clear choices: the voice customers hear and the model responsible for chat and delegated work. You manage both choices in Assistable’s assistant configuration.
Start with a task whose answer depends on information your business already maintains. Service coverage, product details, or appointment availability can make useful examples when the appropriate lookup tool is attached. The tool supplies the business result; the voice delivers it in conversation.
An example: check service coverage during a call
Imagine a home services team with a connected tool that looks up coverage by ZIP code. This illustrative workflow shows how a spoken question can become a specific answer.
- The caller asks about coverage. “Do you provide inspections in my area?” The assistant asks for the ZIP code required by the lookup.
- The assistant checks the connected source. Its configured tools model sends the ZIP code to the service coverage tool.
- The tool returns the answer. In this example, the result says that inspections are available in the caller’s area and lists the applicable inspection types.
- The agent explains the result aloud. It tells the caller which services are available and asks which one they want to discuss.
If the business also has a configured scheduling tool, that conversation can continue into a separate availability or booking step. Coverage alone does not establish an open appointment, and a lookup does not create a booking. The agent should report a completed action only after the relevant tool confirms it.
This is a useful first demo because the team can compare the spoken answer with the returned business data. It also gives reviewers something concrete to assess beyond whether they like the voice.
Choose the model behind chat and tools
The voice drawer includes a Chat backup & tools model setting for GPT Live voices. The assistant header also identifies the Chat & tools model, making the purpose of the selection visible as you configure the assistant.
Use this setting to choose the text model for customer chat and delegated tool work. Selecting a GPT Live voice does not remove that choice or turn the speech model into your chat model.
A saved model change applies to chat immediately. Calls require a full assistant Publish to activate the updated configuration. That distinction matters when your team uses the same assistant across chat and voice: finish saving the configuration, publish it, and test the call experience you intend customers to receive.
Start with one complete customer interaction
A focused rollout makes the experience easier to evaluate. Choose one assistant, one customer question, and one connected tool before adding more complex tasks.
Audition the voices, select the chat and tools model, and review the assistant’s instructions. Make sure the lookup has a clear input and a result the agent can explain. Then use full Publish and place a test call through the intended customer experience.
During that call, listen for how the assistant gathers the required detail, responds to the tool result, and handles a question the tool cannot answer. Check the result itself alongside the spoken response. A good voice preview and a successful business lookup answer different questions; review both before expanding the workflow.
Questions about GPT Live voice agents
Does a voice audition test my connected tools?
No. Auditions play a fixed spoken sample so you can compare voices. A published, tool-enabled assistant call is the place to test your instructions, business data, and actions together.
Do I need to publish after changing a GPT Live voice?
Yes. Use full assistant Publish to activate call changes, including switching to or from GPT Live. Saving a selection or listening to its audition does not by itself update the published call experience.
Can I use GPT Live with a connected Agent Flow?
GPT Live voices cannot currently be used with a connected Agent Flow. If your assistant relies on a flow, keep a compatible voice configuration and review a separate GPT Live setup with our team.
Put your business workflow into the conversation
Bring us one call your team handles often and the business information needed to answer it. We can help assess the voice, connected tools, and assistant configuration for that experience.