Turn interviews
into validity you can show.

A different approach to construct validity.

Construct validity is the degree to which a measurement actually captures the theoretical construct it claims to measure, not something adjacent to it. A survey built to measure job satisfaction should track job satisfaction, not general optimism. A coding scheme built to flag “churn risk” should flag churn risk, not just negative sentiment. Researchers have used construct validity for decades to check that an instrument, a survey, an interview protocol, or a rating scale, behaves the way the underlying concept should.

Why validity erodes between the codebook and the coder

The idea holds up on paper. In practice, validity slips in the gap between a codebook and the person applying it. One rater interprets “urgency” as tone of voice, another interprets it as word choice. A scale built for one study gets reused on a different population without a check. By the time twenty interviews are coded, the earliest ones no longer match the definitions the later ones were coded against, and the construct the team set out to measure has quietly drifted.

Anchoring the construct at every layer

Speak AI applies your rubric, scale, or codebook the same way on interview one and interview two hundred. Each recording is transcribed in your language, with 100+ supported, then scored across three layers: the words, the tone, emotion, and energy in how they were said, and any visuals or screen shares alongside. Every score returns with the exact quote it came from, so a construct score is traceable back to the evidence, not just a number. Scoring can run across multiple models, including Claude, ChatGPT, and Gemini, depending on the task.

What researchers ask their scored data

• “Which interviews score this construct inconsistently across raters, and where does it drift?”
• “Show me every transcript scored below threshold on this construct, with the quote that triggered it.”
• “Does this construct correlate with our outcome measure across the whole dataset?”
• “Which confounding factors show up most in low-confidence scores?”
• “Compare this quarter’s construct scores against last quarter’s.”

From a shaky scale to a defensible dataset

The result is a scale applied the same way across every file, with every score traceable back to a quote instead of a rater’s memory. Scores that once lived in a spreadsheet become dashboards you can customize and white-label, tracking construct scores and drift over time. A legal intelligence firm anchored its own proprietary risk scale into this workflow and processed 5,100+ hours of carrier calls and saved $700K+, scoring every call on the same construct instead of an analyst’s ad hoc read.

And because a valid construct rarely lives alone, the same engine scores calls, interviews, and meetings on the same criteria, connecting construct scoring to call scoring and the broader MCP layer other teams already use.