A caller says, “Friday afternoon works. Actually, can we do Monday morning?” while the agent is still speaking. The agent needs to stop, catch the correction, and avoid confirming Friday by mistake.
On September 10, 2026, OpenAI introduced GPT-Live-1 in the API, describing a full-duplex voice model that can listen and speak simultaneously and delegate reasoning or actions to a backend. That announcement is a useful reason to revisit voice architecture. It is not evidence that a particular business workflow will become reliable simply by changing models.
That booking correction is a useful test case. Follow it through the audio layer, application state, and calendar request, and the engineering work becomes clear.
Separate conversation quality from transaction correctness
A voice agent has two responsibilities: maintaining a useful conversation and interacting correctly with business systems. These responsibilities need different acceptance criteria.
For conversation quality, examine interruptions, thinking pauses, changes of subject, and unclear audio. For transaction correctness, examine whether the agent uses the correct customer record, requests the intended slot, and reports a booking only after confirmation.
Imagine that the caller changes Monday to Tuesday while an availability request is running. The response may arrive after the correction. The application needs a way to distinguish an obsolete request from the current intention. A pleasant voice cannot fix stale application state.
A practical design associates each proposed action with its inputs and a version of the conversation state. Before executing a write, check that the inputs still match the caller’s confirmed request.
Treat interruptions as state changes
Stopping audio playback is only the first part of interruption handling. The system also needs to decide what to do with pending work.
A read-only availability lookup can finish and be ignored if it is no longer relevant. A meeting creation request is different: if it may already have reached the calendar, retrying can create a duplicate. Give write operations a stable request identifier and reconcile uncertain outcomes before repeating them.
Build test scenarios around that distinction:
- A caller interrupts an explanation without changing the task.
- A caller corrects a date before a tool starts.
- A caller corrects a date while a read is pending.
- A caller disconnects after a booking request but before confirmation.
Record the resulting system state as well as the transcript. The agent may sound correct while the calendar is wrong.
Compare architectures against the same workload
A combined speech model and a chained speech-to-text, language-model, and text-to-speech system have different operating tradeoffs. The useful comparison is an end-to-end task on your calls, not an isolated model demonstration.
Use a fixed collection of representative requests. Include names, numbers, noisy lines, impatient callers, and unavailable appointment slots. Measure successful outcomes, correction recovery, time to a useful answer, and the cost of completing the task. Include failed calls and human follow-up in the cost calculation.
Keep a record of provider settings, prompts, and tool definitions for each evaluation. Without that version information, it is difficult to explain why a later release behaves differently.
Start with a bounded deployment
An inbound enquiry workflow is often easier to evaluate than a broad agent that can modify every customer system. Specify what the agent may answer, what it may record, and when it should offer a human next step.
For a property enquiry, the initial outcome might be a confirmed summary of location, budget, and preferred timing. Calendar booking can follow once availability, confirmation, and recovery paths are working. Expanding in stages gives the team a clearer explanation of what changed when results improve or decline.
How this connects to InferScale Voice
InferScale Voice provides visual workflows and browser testing, with phone routes and business-system integrations configured for the deployment. We confirm supported providers and model choices when scoping your environment.
Our voice-agent implementation service can help map a call flow, compare supported components, and test connected actions. Bring one repeatable conversation to a discovery call and define the acceptance criteria before choosing the stack.
Release source: OpenAI’s GPT-Live-1 announcement.