FIGURES · METHOD

WR3 in figures: what we measure, on what sample, and what is missing

By · Published · Updated

WAV Research publishes eight in-house figures. The main one: in a manual verification of the 6 sessions of a 2026 study, WR3 got the speaker right in 98 to 99% of turns, and in 76 to 86% of the passages where several people spoke at once. Separately, an earlier internal measurement on 520 turns from two real 2026 sessions gave 98 to 99% on the doubtful stretches WR3 decides on its own: different figures, on different samples. Here are all eight, each with its sample, its date, the method behind it and what it does not prove.

What are the figures, and on what sample?

This is everything we state with numbers on the site. If a figure is not here, we do not use it.

FigureWhat it measuresSampleDateMethodLimits
98 to 99%Turns that end up with the person who said them when separating who is speaking.6 sessions of a 2026 study, about 12 h, about 67 participants (12 per session on average); manual verification.November 2026Manual verification: a person hired for this transcribed each session by hand, on paper, and we compared turn by turn, listening to the recording, who said each line according to that transcript and according to WR3.Our own measurement, not audited by a third party. The sessions of one wave.
76 to 86%Passages where several people spoke at once that end up with the right person.The same 6 sessions of a 2026 study, about 12 h, about 67 participants; manual verification.November 2026The same manual verification, counted only on the passages with overlapping voices.Our own measurement, not audited by a third party. It only measures the hardest case, when several people speak at once.
98 to 99%Earlier internal measurement: doubtful turns WR3 decides on its own that end up with the right person when separating who is speaking.520 turns from two blocks of one wave: 218 of 223 in one and 294 of 297 in another.October 2026Internal measurement: doubtful turns WR3 decides on its own are compared with a sample of doubtful turns a person labelled by hand.Two blocks of one wave. It covers the doubtful stretches, where each person's own microphone is not enough to decide who is speaking: WR3 decided 21.6 of 37.6 minutes and 19.8 of 31.5 on its own (more than half); a person reviewed the rest.
11 of 120 secondsSeconds in which a lapel microphone's wearer was the loudest voice on their own microphone.One two-minute window from one participant; in another window it was 37 of 120.2026Second-by-second comparison of each voice's level on that microphone.Two windows from one focus group. They show the room-audio problem; they are not an average.
1 in 10Quoted people whose words a hand-built report could attribute to them.One delivered report, built from a single microphone for the whole room.2026Review of the report: how many quoted people had an identified speaker.One report, made without WR3. It shows the limit of a room microphone; it does not measure WR3.
38 of 38Quotes attributed to the right person.The 38 attributed quotes in one delivery of a year-long focus group program.2026Audit of the delivery: each attributed quote was compared with who said it in the recording.One delivery. Not a general error rate.
20+ yearsTrack record of WAV (We Are Vision), the company behind WAV Research.The company's history.As of 2026Company record.Years of experiential marketing, not years of running studies with WR3.
200+ projectsExperiential marketing projects delivered by WAV.The company's project record.As of 2026Company record.Experiential marketing projects, not research studies.

How was each figure measured?

The first six come from real 2026 focus group recordings and deliveries, compared with a reference made by people. The two from the manual verification of a 2026 study are described in how we did it; the detail of each measurement, with its stages and corrections, is in how we measure WR3's figures. The last two are company records: WAV's track record in experiential marketing, which is where the team, the room and the production come from.

One rule keeps us honest: every figure travels with its sample in the same sentence. "98 to 99%" alone says nothing; "98 to 99% in the 6 sessions of a 2026 study, by manual verification" can be checked.

What do these figures not prove?

When are they updated?

From the November 2026 wave we have already published the manual verification of who is speaking. Still to come are the time from recording to delivery and the minutes of human review per session, which we will publish here with their sample and date. This page's update date changes whenever a figure changes.

To see how these figures show up in a deliverable, browse the sample report, built from a fictional study.