VoiceFluency

Retelling

Between reading someone else's words and speaking with no support sits the mode you use most in real life: conveying meaning in your own words — recapping a meeting, an article, a thing that happened. Retelling trains exactly that.

Structure: Read silently → text hides → speak
Delivery score: Same as Monologue
Coverage: Advisory, never in trends
Checked against: The passage's key points

The method

Why it works

The passage disappears before you speak — for good. You rebuild it from meaning rather than memory of the wording, which is what keeps it from quietly becoming a reading exercise.

Memory, not reading
The text hides before the recording starts and does not come back. Whatever you say is reconstructed, not recited.
Structure under load
Holding the through-line while choosing fresh words is the working-memory workout at the heart of clear speaking.
Checked against a source
Unlike free speech, your retelling has a right answer to compare to — so the content feedback names real points instead of giving a vague impression.

How a session runs

Step by step

1
Read the passage silently, at your own pace.
2
The text hides. Start recording and retell it in your own words.
3
You get a delivery score exactly like Monologue's, plus a key-point coverage check.

Evaluation

How it's scored

Retelling is free speech, so it is scored by the Monologue scorer — the identical formula, the identical gates, the identical trend. What it adds is a second, separate verdict about content, which is deliberately kept out of every number.

Deterministic · monologue-deterministic-v2 + advisory coverage

Pace

35%

The same asymmetric Gaussian as Monologue: target 135 WPM in English, 115 in Russian, with a wider slow side.

Cleanliness

40%

The same floored filler and pause rates as Monologue.

Key-point coverage

advisory

Each key point of the passage is marked covered, distorted or missed, each with an exact quote from your transcript as evidence. Reordering clauses is not a distortion — except in process passages, where the order is the content.

The formula

Transcribed from the scorer that ships in the app. Every constant is pinned to this methodology version, so the same take always produces the same number.

Delivery

overall = (0.35 × pace + 0.40 × cleanliness) / 0.75      // identical to Monologue

Coverage

per key point:  COVERED | DISTORTED | MISSED   (+ an exact transcript quote)
plus:  fabricated details · structure kept? · paraphrase quality
       paraphrase = OWN_WORDS | MIXED | VERBATIM

reported as “4 of 6 points” — never mixed into any score

Details worth knowing

Your delivery is the score. Coverage is guidance and never enters a numeric trend — so a hard passage cannot quietly drag down the fluency curve you are actually tracking.
Retelling deliberately belongs to neither the articulation family nor the reading family: it is free speech, so its takes count toward the same fluency trend as Monologue.

When no score is given

A weak guess is worse than no number, so the overall is withheld rather than estimated. What was measured stays visible either way.

Fewer than 10 recognised words, or less than 15 seconds of active speech → no overall, raw metrics preserved.

Reading the number

85–100Excellent

70–84Good

50–69Average

Below 50Needs practice

The coach review

What the AI adds

On top of the usual speech review, the model is given the passage's key points and asked to:

Judge coverage only against that list — it is explicitly forbidden from inventing key points of its own.
Quote the exact words from your transcript that covered or distorted each point.
Flag significant details you added that were not in the source at all.
Say whether you kept the logical structure, and whether you spoke in your own words or fell back on the source's phrasing.

Keep going