How do I improve typing speed with lessons that focus on weak keys?

Use a trainer that measures speed and accuracy for each individual key and then generates practice text weighted toward your slowest letters. TypeCrt's KeyForge does this: it times every keystroke, computes a confidence score per key as targetTime divided by your actual average time, penalised by your error rate, and builds lessons around the lowest-scoring key. This matters because the nine most common English letters make up roughly 70% of typed characters, so ordinary practice text barely exercises the keys that are actually slowing you down.

How does KeyForge decide when to unlock a new letter?

A new letter is added only when every currently unlocked letter reaches a confidence score of 1.0 or higher, meaning you type each of them at or above the target speed. KeyForge starts with six letters (e, n, i, t, r, l) and adds the rest in English frequency order: e n i t r l s o d a c u p g h m b y f v w k x z q j. A newly unlocked letter starts at zero confidence, so it immediately becomes the focused key and dominates the next lessons.

Does KeyForge use real words or made-up ones?

Real words, in roughly 86 to 94 percent of cases depending on how many letters are unlocked. KeyForge draws from a 6,531-word frequency-ranked English corpus built from public domain sources, filtered to words spellable with your currently unlocked letters. Pseudo-words generated from a Markov bigram model are used only as a fallback when the unlocked letter set is too small to spell enough real words.

What is a confidence score in KeyForge?

Confidence is how close a key is to mastered. It equals the target time per keystroke divided by your actual average time for that key, multiplied by an error penalty of one minus half your error rate. The default target is 500 milliseconds per keystroke, roughly 24 words per minute of sustained per-key speed. A score of 1.0 means you have hit the target; below 1.0 means the key is still slow. Scores are smoothed with an exponential moving average so a single bad keystroke does not swing the result.

How is KeyForge different from Keybr?

Both use speed-based per-key confidence and progressive letter unlocking in English frequency order. KeyForge additionally reports a learning rate per key using linear regression over your recent lesson history, so you can see whether a key is actually improving and not just whether it is slow. TypeCrt also carries the practice mode inside a full typing test app with 80 themes, and shows no advertisements, whereas Keybr is ad-supported on its free tier.

Yes. KeyForge is completely free with no account, no signup and no advertisements. Your per-key history is stored locally in your browser.

KeyForge Adaptive Practice

Lessons built from the keys you're actually slow on.

Ordinary typing practice makes you type ordinary English. English letter frequency being what it is, roughly 70% of your keystrokes land on just nine letters — E, T, A, O, I, N, S, H, R — which you almost certainly already type well. The keys costing you speed barely appear.

KeyForge inverts that. It measures every key independently and builds lessons around your worst one. This page documents exactly how, so none of it has to be taken on trust.

Why per-key rather than general practice. The largest published typing study found that specific letter pairs — particularly those typed by different hands or different fingers — predict overall speed better than general measures do, and a motion-capture study found that consistent finger-to-key mapping separates fast typists from slow ones regardless of how many fingers they use. Both point the same way: speed lives in specific transitions, so practice should target them. Figures and citations are on the research page.

Confidence: the core score

Every letter carries a confidence score. It answers one question: are you typing this key at target speed, cleanly?

rawConfidence = targetTime / max(avgLatency, 50ms) confidence = rawConfidence × (1 − errorRate × 0.5)

targetTime — default 500ms per keystroke, about 24 WPM of sustained per-key speed. Adjustable
avgLatency — your mean gap before that key, over a rolling window of the last 30 presses
errorRate — mistakes on that key across the last 30 presses. At most it halves the score, so accuracy matters but never zeroes you out
Minimum data — no score until a key has at least 3 recorded latencies

Scores are then smoothed so one fumbled keystroke can't swing them:

newConfidence = oldConfidence × 0.6 + confidence × 0.4

1.0 is the mastery line. At or above it you're hitting target speed on that key; below it, the key is still costing you.

Progressive unlock

KeyForge starts with six letters and adds more only as you earn them. A new letter appears when every currently unlocked letter reaches confidence ≥ 1.0 — not on a timer, and not on lesson count.

enitrlsodacupghmbyfvwkxzqj

The order is English frequency: e n i t r l s o d a c u p g h m b y f v w k x z q j. Highlighted keys above are the six you begin with.

A newly unlocked letter starts at confidence 0.0, which makes it instantly the weakest key — so it becomes the focused key and dominates the next few lessons. That's intended: new letters get drilled hardest, exactly when they need it.

The focused key

The focused key is simply the lowest-confidence unlocked letter, recalculated after every lesson. Word generation weights toward it, so practice text is dense in whatever is currently worst.

How practice words are generated

Words are filtered to your unlocked letters, then chosen real-word-first:

Unlocked keys	Real words available	Real-word share
4 — `e n i t`	8	~94%
5 — `e n i t r`	14	~93%
6 — `e n i t r l`	27	~92%
7 — `+ a`	83	~91%
9 — `+ o s`	310	~88%
11 — `+ d a c`	821	~86%

The corpus is 6,531 frequency-ranked English words, built from public domain texts, ranked by lowercase-as-written occurrence (which filters proper nouns) and required to appear across multiple independent works (which filters book-specific vocabulary).

With only four letters unlocked you'll see real words repeat — tin, net, tie, inn. That's deliberate. Repeating real words beats inventing fake ones, because muscle memory for real letter sequences is what transfers to real typing.

Pseudo-words are a fallback, not the default. When the unlocked set genuinely can't spell enough English, a Markov bigram model generates pronounceable filler weighted toward the focused key. That path accounts for roughly 6–14% of words depending on stage.

Learning rate

Confidence tells you whether a key is slow. Learning rate tells you whether it's getting better — a different and more useful question.

After each lesson KeyForge stores a per-key sample (hits, misses, average latency), keeping the last 50. Given at least 5 samples it runs a linear regression over the most recent 20 to produce a slope. A negative slope means latency is falling: the key is improving. A flat slope on a low-confidence key means whatever you're doing isn't working, and technique needs to change rather than repetition continuing.

Reading the heatmap

Signal	Meaning	What to do
Confidence ≥ 1.0	At or above target speed	Nothing — it's holding you back no longer
0.7 – 1.0	Close, usually a hesitation not an error	Keep drilling; it will clear on its own
Below 0.7	Genuinely slow or error-prone	Slow down deliberately and fix the finger assignment
Low confidence, flat learning rate	Repetition isn't working	Check you're using the correct finger — usually a wrong-hand or wrong-finger habit
No score	Fewer than 3 samples	Type more; it needs data

Verifying this yourself

Unlock gating — check that no new letter appears while any unlocked key sits below 1.0
Focused key — it should always match the lowest-confidence key on your heatmap
Real-word share — count real words in a generated lesson; it should track the table above for your stage
Error penalty — deliberately mistype one key repeatedly and watch only that key's confidence fall