Undertone interviews a customer by voice, then scores how they answered — pitch range, pace, pauses — against that same speaker’s own baseline, and flags the moments where the words and the delivery disagree.
Below is a real report from a real interview: one participant, twelve turns, every answer playable. Built in a day at the Boson AI Higgs hackathon, 18 September 2026. Source on GitHub.
Eight seconds. The same person, the same sentence, recorded twice.
“Yeah, it was nice. I liked it.”
Identical words, 16 points apart. A transcript cannot tell them apart at all — every text-based research tool records both as “liked it.”
The honest part is where the gap comes from. On pitch range, the axis this engine weights most heavily, the flat take actually scores higher: 17.1 semitones against 12.7. This speaker flattens by slowing down, not by losing pitch. A single axis would have called the flat take the livelier one. That is why every answer is measured on several axes against that same speaker, never against one fixed scale.