Engineering
What real-time voice interviews demand from an AI system
ihiring team · Aug 22, 2026 · 2 min read
An AI interviewer can look simple from the outside: a candidate talks, the system answers. Making that feel like a conversation, rather than a phone menu, is a real-time engineering problem.
Latency is the product
People notice delays in conversation within a few hundred milliseconds. A long pause after every answer makes an interview feel mechanical, and it makes candidates more anxious.
So the system is built as a streaming pipeline, not a chain of requests and responses. Audio is transcribed as the candidate speaks, the next question is prepared while they finish, and the reply is spoken as soon as it is ready.
Knowing when the candidate has finished
Turn-taking is harder than it sounds. A pause might mean the candidate is thinking, not done. Cutting someone off is worse than waiting a moment too long, so the system weighs silence together with the content of the answer.
It also has to handle interruptions gracefully, such as a candidate asking for a question to be repeated.
Audio and video that survive real networks
Candidates join from phones, laptops and shared Wi-Fi. Live sessions use WebRTC, which adapts to weak connections, and the interview has to degrade gracefully: keeping the audio clear even when video quality drops.
Scale without queueing
Every interview is a live session, so capacity matters. Sessions run in parallel, which means a surge of applicants does not become a waiting list. The challenge is keeping the experience the same whether ten or a thousand interviews are running at once.
A transcript you can trust
Transcription quality affects everything downstream: follow-up questions, summaries and reports. Accents, background noise and technical vocabulary all matter. The transcript is stored next to the video, so your team can check any summary against what the candidate actually said.