The engine room

How Shadowin actually works

Shadowin is built on one simple loop: you hear a native-quality sentence, you say it back out loud, and a real speech engine grades how close you got — word by word, in color. No vague stars, no black box. This page opens the hood and shows you exactly what every part of the system is doing, so you can practice with intent and squeeze the most out of every minute.

Start practicing See Premium

Core mechanics

The four-step shadowing loop

Every attempt runs the same path. It takes a couple of seconds, and the same engine powers your very first onboarding baseline and every lesson after it.

01
Listen
Hear the line at native pace, as many times as you like. Shadowing is about matching rhythm, not just words.
02
Record
Say it back. Your audio is captured in the browser and sent securely to the scoring service — never stored beyond the moment it is graded.
03
Align & score
The engine lines the target line up against your audio and scores every sound inside every word — not just whether the word arrived.
04
See it in color
Each word lights up green, amber, or red, with a single accuracy score from 0–100 — and red points at a sound you personally keep missing.

Reading your results

What the colors mean

The sentence is not marked against a fixed pass mark. Green, amber and red are three different statements, and red is the one that is about you specifically.

Green — nothing to fix here

The word scored 75 or above and none of your problem sounds slipped in it. Example: The future is

Amber — imperfect, but not your problem

Something in the word came out under the bar, but none of it is a sound you routinely miss. Worth a glance, not worth a retake. Most of what used to go red now lands here.

Red — one of YOUR sounds slipped

Red is personal. A word turns red only when it contains one of the five sounds you miss most often — measured across every take you have ever recorded — and the engine flagged that exact sound in this one. Red never means “low score”. It means here is the thing you keep getting wrong, and these are the words to loop before you retry.

Why the chips below disagree

The per-word strip under the score answers a different question: it is the engine’s raw verdict on each word, with no idea who you are. So a word can be red above (your sound slipped) and green below (the word as a whole still passed) — or the reverse. Above is about you; below is about the take.

Where your five sounds come from

Every sound you produce is scored and kept. Once a sound has at least 40 attempts behind it, it is ranked by how often it slips — not by its average score, because a sound can hold a respectable average and still fail one take in three. The worst five are the ones that turn words red. The list is yours alone, and it moves as you improve: fix a sound and it drops out, and the next one takes its place.

Under the hood

The grading engine, demystified

A pronunciation grader is only as good as how honestly it listens. We made deliberate engineering choices so your score reflects how you really spoke — never inflated, never unfairly harsh.

It scores sounds, not words

The engine already knows the line you were given, so it does not have to guess what you said. It lines the reference up against your audio and grades every individual sound inside every word. That is why the feedback can name which sound slipped instead of just marking the word wrong.

It builds a profile of you

Every scored sound goes into your own phoneme profile — how often each one slips, and what the engine hears instead. That profile is what makes red personal, and it is what the sounds strip on your dashboard is drawn from.

Two scorers, never averaged

When the engine cannot answer a take, a backup scorer grades it instead — and it grades on a different scale. The two are never blended into one number. The pass marks move with the scorer, and a take graded by the backup says so on the result screen.

Mumbling can't game it

A sound that never really arrived scores as missing rather than passing quietly. Speak clearly and you get full marks; swallow half the word and the sounds inside it are what pay for it.

Things we refuse to penalize you for

Before comparing, both texts pass through smart normalization, so harmless variations always count as a perfect match.

Contractions

“do not” = “don’t”. Say it either way.

Numbers

“30” = “thirty” (up to ninety-nine).

Currency

“$30” reads as “thirty dollars”.

Compounds & hyphens

“lifecycles” = “life cycles” = “life-cycles”; “setup” = “set up”.

Smart quotes

Curly and straight apostrophes are treated the same.

Name variants

Common spellings of the same name are clustered together.

Lesson tiers

Three tiers, drawn by word count

Every lesson is sorted into one of three tiers — and the tier is decided by a hard rule: the number of words in the sentence. The badge you see always matches the filter you picked, with zero AI guesswork.

A1–B1

Standard

5–10 words

Single, natural sentences in everyday conversational English. Concrete and practical — the things real people say out loud every day. The place to build clean fundamentals.

B2–C1

Advanced

11–18 words

Richer, multi-clause lines in a confident professional and social register. Natural idioms and real, specific points — never stiff or academic. Where fluency starts to feel effortless.

C1–C2

Executive Masterclass

19–35 words

Long, idea-dense sentences at native pace, written in the cadence of a sharp boardroom or a long-form intellectual podcast. No robotic textbook phrasing — just elite spoken English.

Why word count, not vibes? The tier badge and the library filter are bound to the same measurement at the moment a lesson is created. That means an “Advanced” lesson can never quietly show up under the “Standard” filter — what you filter for is exactly what you get.

One bar, the same for every lesson

There are no strictness modes to choose. Every attempt is graded the same way, so a score means the same thing today, next month, and for the person practising next to you.

What earns XP

Two things, both from the same recording: your score clears the pass mark, and most of your sounds landed. A high number carried by a few perfect words while a quarter of the sounds slipped does not count.

When it doesn’t

You are told why. If the score looks good and no XP arrived, a line under it says how many of your sounds did not land and how many are allowed — so the next take has a target, not a mystery.

Beyond the number

Your four mastery dimensions

One score tells you how close you were. The mastery breakdown tells you why — four angles on the same attempt. It is derived from a word-by-word transcript, so it appears on takes graded by the backup scorer rather than on every take.

Cadence

Rewards long, unbroken runs of correct words — your rhythm and flow.

Accent

Drops when misses are scattered across the line (drift) rather than clustered in one spot.

Phrasing

Penalizes your single longest stumble — one tripped phrase reads differently from a few stray slips.

Dynamic range

A composite of flow and consistency that rewards expressive, confident delivery.

The AI Coach nudge

After each attempt, Shadowin finds your weakest of the four dimensions and hands you one specific, actionable cue — slow down and exaggerate your vowels, loop a rough patch in your head before retrying, vary your intensity. Small, targeted, and different depending on what tripped you up.

Prefer a calmer screen? Minimalist Mode in Settings hides this breakdown and shows just your score and word colors.

Personalization

How your daily practice is built

Shadowin doesn’t hand everyone the same feed. What you see is shaped by the topics you chose, what your review schedule says is ripe, and what you have already practised.

You choose your priority topics

In Settings and on your Dashboard you pick the domains you care about — Business, Tech & AI, Philosophy, Finance, Longevity, and more. Those choices float matching lessons to the top of your library stream.

Replay brings back what you have done

One button on your dashboard reshuffles lessons you have already practised into a fresh session. Free accounts get one session of ten lessons a day; Premium gets five sessions of fifteen. There is no timer inside a replay — only the session count.

What decides the order you see lessons in

1
Due for review
Lessons your spaced-repetition schedule says are ripe today — caught right before you’d forget them.
2
Your priority topics
The domains you picked in Settings float matching lessons to the top of your library stream.
3
Your own history
Replay draws only from lessons you have already finished, shuffled so the order never becomes a script you memorise.

Staying in the game

XP, levels, streaks & your daily window

The systems that keep momentum honest — generous enough to reward real practice, structured enough that progress means something.

XP & levels

Clear your mode’s bar and bank XP. Levels follow a rising curve — the early ones come fast, then each one asks a little more, so a high level genuinely reflects the work behind it. Premium earns XP 1.5× faster.

Streaks

Practice on consecutive days and your streak climbs. Miss a day and it resets to one — counted in UTC so it’s consistent across every device you use.

Free daily window

Free accounts get 10 minutes of regular practice per day, enforced on the server. Premium removes the cap entirely for unlimited daily reps.

Replay sessions

A shuffled chain of lessons you have already recorded — and exempt from the 10-minute window, so a second pass is always available.

Your voice, your data

What happens to your recordings

Your recording goes to our own pronunciation engine, which scores it and keeps only the numbers — not the audio, unless you say otherwise. Each attempt is stored as a row of per-sound scores and the recording is discarded once it has been graded. If you want to help the engine get better you can switch keeping your recordings on in Settings — it is off until you do, and switching it back off stops new recordings being kept. OpenAI is involved only when our engine cannot answer a take and a backup transcription is needed, and under their API terms those inputs are not used to train their models. What we keep is the non-identifying result that powers your progress — your scores, the sounds you miss most, your streak. Full details live in our Privacy Policy.

Now you know the machine. Go feed it your voice.

Every word you read out loud sharpens the same engine that grades you. Pick a topic, hit record, and watch the colors turn green.

Start a session

Want something specific? Ask for it.

Missing a topic, a word you keep tripping on, or an idiom you want to nail? Tell us and we’ll add real practice material for it.

Want Andrii to personally add your favorite topics, words, or phrases? Sign in to submit a request!

Listen

Record

Align & score

See it in color