Voice AI screening
Voice AI screening is a real-time spoken conversation between a job candidate and an AI agent that asks job-relevant questions, follows up adaptively on the candidate’s answers, and writes a transcript scored against a recruiter rubric.
Direct answer
Voice AI screening is a live two-way phone-or-browser call where one side is an AI agent and the other is a human candidate. The agent listens, decides what to ask next, and speaks the question — the same loop a recruiter runs, only typed out by software and pinned to a rubric. It matters because it is the cheapest way to give every applicant a fair first-round when the recruiter does not have the hours. Who needs it: teams hiring more candidates per week than there are first-round slots on the recruiter's calendar.
In detail
The pipeline
Three components run in parallel. (1) Speech-to-text (STT) transcribes the candidate audio with a streaming model — Deepgram, AssemblyAI, Whisper and similar. (2) A large language model receives the running transcript plus the system prompt (role brief, rubric, time cap) and decides whether to ask the next planned question, follow up on the last answer, or wrap. (3) Text-to-speech (TTS) — Cartesia, ElevenLabs, etc. — renders the chosen response back into speech. WebRTC carries the audio in both directions. End-to-end latency under one second is the bar for the conversation to feel natural.
When it applies
Voice AI fits screens where signal comes from how a candidate explains themselves — backstory, motivation, ownership, communication. It fits less well where the screen is actually a skills test that wants typed code, a portfolio review, or a controlled assessment environment. For those, voice AI is upstream (qualify intent and basic fit), then a human or async task takes the deeper look.
Common misconceptions
It is not one-way video — there is no recording of the candidate talking to a camera with no listener. It is not an IVR phone tree — there is no fixed menu, the agent forms sentences in context. It is not proctoring — well-designed voice AI does not run gaze tracking, screen capture, or cheat-detection overlays. And it is not the same product as an outbound recruiting bot that calls candidates to pitch a role; the surface looks similar but the intent is opposite.
Related concepts
Voice AI screening is the conversation surface for an AI first-round interview. When the same questions and rubric run every time, it qualifies as a structured interview. The score uses rubric scoring. In NYC, if that score weighs in on the hire decision, the tool is also an AEDT.
How Vettika handles it
Vettika runs the conversation on LiveKit with Deepgram STT, a Gemini-class LLM for turn planning, and Cartesia TTS. Twelve-minute hard cap. Audio is destroyed on hangup by default; the transcript is what gets kept. See the voice agent product page.
Try a first-round interview, no card.
90-second sample available from any page. No signup, no email.
Run your first 3 free