The anatomy of a voice interview: what Ploras actually listens for

Under the hood of the Voice Agent — from prosody signals to fraud vectors. A visual walkthrough of every scoring dimension.

"Voice AI" is a category label. What actually happens in a Ploras screen is closer to seven simultaneous models running on the same audio stream. Here's what each one is doing.

The seven dimensions we score

Clarity & prosody

Not "accent" — communication cadence, pause structure, filler rate. Weak correlation with hire quality on its own; strong when combined with the others.

Domain depth

Fine-tuned prompts probe actual know-how. Candidate explains a specific tradeoff; model scores whether the explanation matches someone who's actually done the work.

Fraud signals

Voice vs. ID match. Second-voice detection (is someone else answering?). Copy-paste response detection. If you can automate it, someone will try to — we watch for it.

The interesting signal isn't in any single dimension. It's in how the dimensions move together. Strong domain depth + weak clarity often means a senior engineer who's communication-rusty — a hirable signal most scoring rubrics miss.

Design principle:: Every score we expose has a sentence-level explanation attached. If a human reviewer can't read why a candidate got a 7.2, we don't ship that score.