How Exprea Scores a Speaker

By Noam Shemla · 15 February 2026

By Noam Shemla · Data & Method

TL;DR A Speaker Signature is five scores from 0 to 100: Clarity, Tonality, Structure, Energy and Intellect. It is a weighted average of everything you have ever done, where each source counts differently on each axis and older evidence fades on a sixty-day half-life. Your last recording is only one input.

Most speaking feedback is a person's impression of you, delivered with confidence. It is often useful and almost never checkable. The alternative most products offer is a number, which looks checkable and usually is not. A score with no stated instrument is an impression wearing a lab coat.

This article is the instrument. It describes exactly how a Speaker Signature is produced, which parts of it are measurement and which are judgement, and where the method is weaker than the interface implies.

The five axes

Clarity is how easy the message is to follow. Tonality is vocal variety: how much the voice moves, whatever its height. Structure is whether the thing you said was built. Energy is audible commitment. Intellect is the depth and lexical richness of the ideas themselves.

The shape matters more than the area. A speaker at 85 on every axis and a speaker at 85 on four and 40 on one have very different problems, and averaging them into a single "speaking score" would hide the only fact worth acting on.

What a Signature averages

This is the part most easily misread, so it comes before the mechanics.

Every recorded practice does produce its own five scores. The Signature is not those scores. It is a weighted mean across every activity you have logged (recorded practices, live sessions, Live Battle turns, and the votes other members cast on you), and each of those contributes through three multipliers stacked together:

Three multipliers, stacked Every contribution to every axis is multiplied by all three before it is averaged in. They are independent of each other. A perfectly-measured recording from eight months ago and a peer vote from yesterday can end up carrying similar weight, for entirely different reasons.

The consequence worth internalising: one bad recording cannot crater your Signature, and one excellent one cannot fix it. It moves like a running average, because it is one.

Two analysis paths

There are two ways a recording gets analysed, and which one runs depends on what is configured.

The accurate path. Audio goes to Deepgram's nova-3 model, which returns every word with a start time, an end time, and a confidence. If the average word confidence clears 0.55, those timings become a metrics object: words per minute, a pace timeline in ten-second buckets, the position of every filler word, the location and length of every pause. That object, plus the transcript, then goes to a language model, which writes the coaching and assigns the five scores.

The estimated path. Otherwise the raw audio goes to Gemini instead and the model scores it by ear. No timing measurements are taken at all.

"Otherwise" is doing more work than it looks. The measured path needs two separate services configured: the transcriber and the model that reads its output. If only the transcriber is available, the transcription still runs and is then discarded, and you get the estimated report anyway. The path also falls back when the audio is noisy enough to drop mean word confidence below 0.55, which means the quality of your microphone decides which of the two methods scores you, and nothing in the report tells you which one ran.

Deepgram nova-3, word-level timestamps with per-word confidence; In-browser Web Audio analysis, 2048-sample window, sampled 4× per second of recording

Measured

• Words per minute, overall and per 10-second bucket
• Every pause longer than 0.5 seconds, with its length
• Every filler word, with its timestamp
• Pitch mean and standard deviation, in hertz
• Loudness average, minimum and maximum

Inferred, not measured

• All five axis scores, 0–100
• The overall score and grade
• All written coaching

What the instrument sees

This is the shape of the data a single practice produces, the same structure the model is handed.

Pace is measured, and deliberately not an axis

Words per minute is the single most precisely measured thing in the entire system. It is also, deliberately, not one of the five axes.

The reasoning is that pace on its own is not a quality. Nobody has ever been a good speaker because they spoke at 140 words per minute. Pace only matters through what it does to other things: a rushed delivery is hard to follow, a flat one sounds disengaged. So the measurement is kept and the axis is dropped, and the number is decomposed into the qualities it affects.

That value is then split across three axes and never scored on its own:

This decomposition, one measurement split at different weights across the axes it informs, is the mechanism the whole system is built on. Everything else is an application of it.

Four sources, and what each is allowed to say

A Signature draws on four kinds of evidence, and they are not equally believable. They also differ per axis: a transcript from a live session is decent evidence about your filler rate and useless evidence about your vocal range, because it has no audio in it at all.

So trust is set per source and per axis, as a grid:

An n/a cell means that source is structurally incapable of speaking to that axis, so it is left out of that average instead of being given a cautious middle value. A session transcript is never quietly counted as evidence of your Tonality at low confidence. It is left out entirely.

Two rows are worth reading twice. Live Battle turns score higher on Structure and Energy than an ordinary recording does, for both speakers and whatever the result. A timed, real-time, adversarial turn is simply a more telling sample of those two qualities than an untimed solo recording, and gating it on the outcome would reward beating a weak opponent instead of performing under pressure.

Peer votes are the only source with a human on the other end, and they are trusted at 0.4 across the three axes an audience can perceive. They are given no say on Clarity or Intellect, because a listener's impression of how articulate someone was is exactly the unmeasured intuition this system exists to supplement.

Which parts are computed and which are judged

Clarity is the one axis with a published formula, and it runs alongside the model rather than instead of it. The formula starts at 100 and subtracts six points for every one percent of your words that is a filler, three for every one percent that is a hedge:

Every other axis is a language model reading measurements it did not take, instructed that an average clip lands between 50 and 65 and that 85 and above is reserved for excellent delivery. Intellect is the partial exception: alongside the model's judgement, a type-token ratio (unique words divided by total words) is computed from the transcript and contributes at 0.6 weight.

These numbers are more grounded than a human impression, because the model is handed the statistics and cannot claim you were monotone when the measured pitch deviation says otherwise. Grounded as they are, they remain judgements.

A caveat about Energy Energy is not a physical unit. It is a normalised loudness figure derived from your microphone, scaled and clipped to a 0–1 range. Microphone gain, automatic level control, and how far you sat from the laptop all move it. Two identical performances on two different machines will not produce the same Energy number, and no calibration step corrects for this.

Nothing expires, it just fades

There is no window. An earlier version of this system counted only the last ninety days, which meant a recording silently stopped existing on its ninety-first. That cliff made the number jump for reasons that had nothing to do with you.

Every activity now counts forever, weighted by an exponential decay with a sixty-day half-life:

The thirty-word floor

An activity contributes nothing at all if it is known to be shorter than thirty words.

This exists because of a specific failure. A twelve-word abandoned Live Battle turn used to contribute the model's five axis scores at full weight and a computed clarity of zero, because one filler in twelve words is an eight percent filler rate. A single walked-away-from turn visibly dented a member's radar.

An activity whose length is simply unknown is allowed through. The floor is there to reject known junk. Demanding that every historical record prove its length would delete real history from older recordings that never stored a word count.

Where the product contradicts itself

Five different "ideal pace" ranges ship in the app at the same time. The function that converts your wpm into a score uses 130–160, and the prompt that writes your coaching agrees with it. The pace-focused coaching track says 120–150. The milestone that congratulates you for hitting an ideal pace uses 120–160. The resource recommender treats 110–160 as acceptable and tells you to aim for 130. The gauge you watch while recording highlights 120–150.

The scoring function is authoritative, because it is the thing that produces the number. The others are coaching heuristics that were tuned separately and never reconciled. Every article in this collection quotes 130–160 for that reason.

There is a second inconsistency, and it is stranger than a disagreement between two paths: two filler-word lists run on the same recording, and both numbers are shown to you. A sixteen-pattern list produces the filler count and the timestamps your coaching text points at. A separate twenty-four-pattern list produces the filler rate that gets stored, trended and used in the community benchmark; it additionally counts well, okay, right, yeah, honestly, obviously, anyway and to be honest.

The count and the percentage on a single report are therefore not arithmetically consistent: you cannot multiply one by your word count and arrive at the other. The percentage, being drawn from the wider list, should be read as an upper bound.

A third list of eight items appears in the prompt that describes fillers to the model on the estimated path, which is a third answer again to the question of what counts.

Three limits of the score

The score does not rank you against other speakers. The community benchmark, where it appears, compares your last ten practices against everyone else's ninety-day average. The windows differ, so read it as a rough position and not a league table. It is suppressed entirely below three members, because a percentile drawn from two people is theatre.

Nor is it evenly supported across the five axes. A live session, which for many members is the majority of their speaking, can only contribute to Clarity and Intellect: there is no audio to measure a voice from and no model reading of the argument. Your Tonality and Energy come almost entirely from recorded practice and from what other members voted, and if you do neither, those axes are thin.

Finally, the score exists to make a change visible: filler down from eight percent to five, pace steadied through the middle two minutes. A number that never moves is not measuring anything you did.

Go deeper: What Is an AI Speech Coach, and What Can It Do?