What a speaking score can and cannot tell you

A speaking score is a useful mirror and a bad verdict. The difference matters, because a number invites you to treat it as a judgement on you rather than a measurement of one recording.

Here is what the number actually contains, and where it stops being useful.

What it measures well

Fillers. A count of “um”, “uh” and friends, from word-level timestamps, divided by minutes of actual speech. This is close to objective. Unless the transcription misses a word, the count is the count.

Pace. Words per minute, fillers excluded. Also close to objective, and easy to verify: count the words, divide by the length of the speech.

Long pauses. Silences longer than 1.5 seconds between words. Timestamps again, so this is reproducible.

Hedging. A fixed list of phrases — “I think”, “maybe”, “just”, “kind of”, “sort of” — counted per minute. Reproducible, but a blunt instrument: it cannot tell a genuine uncertainty from a verbal habit.

What it measures approximately

Structure. A language model reads the transcript and judges whether the point came first, whether it was supported, whether it closed. Two competent listeners might disagree by ten or fifteen points. The score is a consistent opinion, not a measurement.

Conciseness. Same mechanism, same caveat. A dense 90-second answer and a padded 60-second answer can score similarly, because the model is reading text rather than hearing the room.

The score weights the objective parts more heavily — fillers are 30% and pace 20%, while structure and conciseness together are 25% — which is deliberate. When the model fails, those parts are dropped and the remaining weights are rescaled; the number still means something, it just means less.

What it cannot tell you

• Whether the idea was good. A clear, well-paced, empty answer scores well.
• Whether the room was hostile. A nervous answer to an unfair question can score low and still be the right thing to have said.
• Whether you were understood. Comprehension depends on shared context, jargon, audio quality and the listener's attention — none of which are in the transcript.
• Whether you sound like yourself. The score rewards a particular kind of clean, structured speech. That is a useful default at work, not the only good way to talk.

How to use it without being ruled by it

Compare yourself to yourself. A 62 on a hard question after a 58 last week is progress; a 90 on an easy one tells you less. This is why the product shows history and component scores rather than a single leaderboard number.

Look at the component that moved, not the total. The total can stay flat while fillers drop and pace gets worse. The component breakdown is where the next drill comes from.

Treat a single result as a sample. Three results is a pattern; ten is a trend. Anything less is a snapshot of one morning.

The honest summary

The number is worth having because it is consistent and it is yours. It is not a measure of your ability, your English, or your worth in a meeting. It is a measure of one recording, on one day, against a public formula — and that is enough to find the next thing to practise.

If you want to see the formula itself rather than a marketing summary, it is written out on the how it works page, with the thresholds and weights.