Google
S08
HardPremiumDesign a Real-Time Voice Agent
Design a phone agent that books and changes appointments in real time: hears the caller, decides, speaks, and handles interruptions and slow tools.
StreamingLatency budgetTurn-takingBarge-inSTT/TTSTool use
Key Requirements
- First agent audio within about 1 s of the caller finishing, at p95, measured per stage (assumed budget)
- Barge-in: stop speaking at once when the caller talks; ignore mm-hm backchannels
- Read back and confirm any booking or cancellation before it is written
- Work on narrowband phone audio with noise and accents
- Hand off to a person with the transcript when stuck or asked
Interviewer Follow-ups
- Q1How do you handle the user interrupting mid-sentence?
- Q2What happens when the speech model misrecognizes a key word?
- Q3How do you manage the latency budget across STT, LLM, and TTS?
Loading...