Live calls: interruptions, changes of mind and hand-off to a person
What the agent does when a caller talks over it, changes a booking or asks for a human, and what is still not finished.
The three things that break live calls
People who build call agents say the demo falls apart in the same places: the caller talks over the agent, changes their mind halfway, or asks for a human and has to repeat everything. These are the behaviours Speakvora's agent was changed to handle in version 0.40.0, and they are all in the downloadable agent today.
What is handled, and how we know
- Interruptions: the agent keeps a record per call of what it said and how much of it was heard, so the next answer does not assume the caller heard everything.
- Corrections: a changed time or detail replaces the old one in the booking, the calendar entry and the summary. One booking, never two.
- Hand-off with details: the person who takes over gets the issue, the corrected details, the confirmed actions and the pending work.
- Honest summaries: built from what actually happened, so a refused booking is never reported as booked.
- No duplicates on retry: repeated requests do not repeat a text or a booking.
The test we used: a caller starts a booking, changes their mind, then asks for a person; the person must receive the corrected details and the summary must say exactly what was and was not booked. It passes through the local agent and through the hosted API route in our automated tests. Real phone calls were run on the local agent with a real Twilio number; the full list of what we tested and did not is on the update page and in the agent's documentation.
What is not done yet
No latency numbers are published yet, streaming audio is still sentence by sentence, an unanswered transfer creates a callback task and a team notice (hold music and voicemail are next), and there is no live coaching mode for human reps. We would rather say so than imply it.
Try it
Start with the AI front desk for small business, see the free course for each step on a real dashboard, or read test your voice agent before going live. You pay once per business location. Speakvora charges no licence fee per call, per text, per minute or per conversation turn; your phone provider bills your calls and texts, and an AI key is optional and billed by your AI provider. New lines the agent needs are voiced once through the Speakvora API at $3 per million characters, and a line you save to your library replays at no further charge; speech to text is $0.0015 a minute. Pricing.
Questions
What does the agent do when a caller interrupts it?
It records which part of its answer the caller actually heard and keeps that separate from what it meant to say. On a phone call through Twilio the heard part is an estimate (Twilio reports partial speech, not playback), and the agent flags it as such, for example heard about 60 percent. Stopping a sentence never counts as the caller cancelling anything.
What happens if the caller changes their mind in the middle of a booking?
The booking is changed, not duplicated: there is one booking, the new time or details replace the old everywhere, and the correction history is kept. A request that was refused is never reported as booked.
Can the caller ask for a person, and does the person have to start over?
Yes. When a caller asks for a person (or the agent is stuck), the agent passes the call with a short brief: the issue, the corrected details, what is confirmed and what is still pending. On the local agent the person hears the brief whispered before the caller is connected and a notice goes to Slack or Teams. The hosted speech API carries the same brief in its hand-off event.
Does the summary say what was booked or only what was discussed?
It is built from outcomes: confirmed, changed, cancelled or not booked, then the open items. Nothing is listed as booked unless the booking really happened.
Will a retry send the same text or make the same booking twice?
Not on the Twilio path: a retried request is recognised by its idempotency token and not repeated.
What is not finished?
We have not published latency benchmarks or compared ourselves with other platforms. Streaming audio beyond sentence-by-sentence, hold music and voicemail for an unanswered transfer, a live assist mode that coaches a human during a call, and a dashboard view for follow-up tasks are planned, not shipped. The voices are AI-generated, and there is no HIPAA claim.