Swipe Typing

Swipe typing lets you type words by drawing a continuous path through letters. CleverKeys ships two on-device prediction engines — a CTC model (the default) and a geometric decoder — and the Prediction Engine setting chooses between them.

Quick Summary

What	Description
Purpose	Type words faster by swiping
Gesture	Draw path through letters without lifting finger
Engines	CTC (default) and geometric — selectable in Settings

How It Works

Instead of tapping each letter:

1. Touch the first letter of your word
2. Slide your finger through each letter in order
3. Lift your finger at the last letter
4. The word appears in the text

For example, to type “hello”:

• Touch h → slide to e → slide to l → slide to l → lift at o

How to Use

Step 1: Start on the First Letter

Touch and hold the first letter of your word. A trail appears showing your path.

Step 2: Draw Through Letters

Without lifting your finger, slide through each letter in sequence. You don’t need to be perfectly accurate - the AI predicts your intended word.

Step 3: Lift to Complete

Lift your finger when you reach the last letter. The predicted word appears in your text.

Step 4: Choose from Predictions

If the wrong word appears:

• Check the prediction bar for alternatives
• Tap the correct word to replace it

Tips for Better Accuracy

• Start and end precisely: Begin and end on the correct letters
• Hit key letters: Pass through distinctive letters in the word
• Maintain steady speed: Don’t rush or pause mid-word
• Use longer swipes: Longer words are easier to predict than short ones

[!TIP] Double letters (like ‘ll’ in ‘hello’) can be swiped in a small loop or just passed through once.

Prediction Bar

After swiping, predictions appear in a horizontal row with the best match on the left. The first suggestion is inserted when your app accepts the word. If insertion fails, the candidates stay available and the keyboard does not treat the word as inserted.

Tap any prediction to use it instead.

Choosing a Prediction Engine

The Prediction Engine dropdown (Settings > Swipe Typing) selects which decoder handles your swipes. The right engine is picked automatically per swipe, based on your layout and language:

Mode	Latin layout, CTC-served language	Non-Latin layout with a wired script (Cyrillic, Greek, Hebrew)	Anything else
CTC (default)	CTC	CTC (per-script encoder)	Geometric
Geometric	Geometric	Geometric	Geometric

• CTC — the CleverKeys-trained decoder. In our benchmark on 2,400 held-out English swipes it got the intended word right on the first try about 89% of the time. It serves three groups:

  1. The 7 bundled Latin languages — English, French, German, Spanish, Italian, Portuguese, Swedish — on any Latin layout that has all 26 letters (QWERTY, AZERTY, QWERTZ, Dvorak, Colemak…).
  2. 6 non-Latin languages via script routing — Russian, Ukrainian, Bulgarian and Macedonian (Cyrillic), Greek, and Hebrew. Each has its own layout, emission alphabet and encoder, and becomes available once you import that language’s pack.
  3. Any imported Latin language pack whose vocabulary is a–z-typeable. CleverKeys measures the pack itself rather than trusting a label: Dutch, Indonesian, Malay, Tagalog and Swahili all measure 100% typeable and are served.

  Turkish is the deliberate exception. Dotless ı has no a–z spelling at all, so a quarter of the Turkish vocabulary — and a sixth of its thousand most common words (nasıl, artık, mı, aynı) — could not be swiped under CTC. Turkish is routed to the geometric engine, which decodes over the board’s real keys and can therefore reach those words. The same test rejects Polish ł, Vietnamese đ and Icelandic þ/ð packs.

• Geometric — a layout-agnostic shape-matching decoder. It is the automatic fallback when CTC prerequisites are unavailable. Both engines require a usable dictionary, supported letter geometry and a valid gesture. Existing Bangla tap layouts do not yet provide Bangla prediction or swipe support.

How much we actually know per language

Only English, French, German and Spanish have their own measured accuracy bar; the 89% figure above is English.

Italian, Portuguese and Swedish were added on 2026-08-18 and are provisional. We have never been able to measure CTC’s accuracy on them — no public swipe corpus exists — so what carries them is the decoder setting they share with French, German and Spanish (it is calibrated against the dictionary’s frequency scale, not the language) plus the fact that the model reads key positions and never sees a language at all. The alternative was the geometric engine, which has no measurement on those languages either and lost by 15–22 points everywhere both were measured.

Russian is validation-tier: it has a real-swipe probe, but not a held-out test bar of the kind English has.

Greek, Ukrainian, Bulgarian, Macedonian and Hebrew, and every imported pack, have no real-swipe probe at any tier. An imported pack is provisional permanently and by construction — the word list is a file you brought, so there is no corpus to measure it against. No accuracy number is published for any of these, and any figure you see quoted for English/French/German/Spanish does not apply to them.

Accented words work normally: a swipe traces the unaccented letters (there is no separate “é” key on the path), and the engine inserts the dictionary’s accented spelling — swipe c-a-f-e in French and you get “café”. Where two words share the same unaccented spelling, the more common one is offered.

[!NOTE] Before v1.6.0 there were two further modes, Neural and Hybrid, backed by an ONNX transformer. That engine was removed: CTC beat it by a wide margin on the same test set (89% vs 75% first-try) while the transformer only worked on QWERTY and cost about 10 MB of app size. If your device still has “Neural” or “Hybrid” stored, it now behaves as CTC.

Whichever engine decodes a swipe, the results flow through the same suggestion bar, autocorrect, and contraction handling (“dont” is shown as “don’t”). If you enable suggestion origin markers, each suggestion is tagged with the engine that actually produced it.

CTC Settings

With the CTC engine selected, a Full CTC Settings button appears with one tuning knob:

Setting	Default	Range	Description
Beam Width	100	10–300	How many word hypotheses the decoder keeps while tracing your swipe. 100 is the validated default; higher values cost CPU per swipe for marginal accuracy.

The CTC scoring constants are calibrated offline and deliberately not user-tunable.

Settings

Tune swipe typing in Settings > Swipe Typing:

Setting	Description
Swipe Typing	Enable/disable swipe input
Prediction Engine	CTC / Geometric (see above)
Swipe on Password Fields	Allow swipe typing in password fields (default: off)
Backspace Undo Swipe	Backspace deletes entire swiped word + auto-space (default: on)

Undoing a Swipe

If the wrong word was predicted after swiping:

Quick Undo with Backspace

Press backspace immediately after swiping (before typing anything else). The entire swiped word and its trailing auto-space are deleted in one press, so you can swipe again.

This behavior is controlled by the Backspace Undo Swipe toggle in Settings > Word Prediction (enabled by default). When disabled, backspace deletes a single character as normal.

Choose an Alternative

Check the prediction bar — alternative words are shown left-to-right by confidence. Tap any alternative to replace the auto-inserted word.

Prefer a Word When Swiping

If the keyboard keeps reading one of your swipes as the wrong word — you swipe “git” and get “got” — correct it the usual way (tap the intended word in the bar, or backspace the swiped word and type or swipe the one you meant). After you have made the same correction twice, the bar asks “Prefer “git” when swiping?” with a Don’t ask option next to it:

• Accept adds the word to your personal dictionary, which makes the swipe decoder favour it from then on (“Swiping now prefers “git””). Tap that confirmation twice to undo it.
• Don’t ask remembers your answer for that word.

If the word is already in your personal dictionary and you keep correcting swipes to it, the bar asks again after two more corrections. Accepting then raises the word’s swipe priority to High (“Swiping now strongly prefers “git””). Tapping the confirmation twice puts the word back to Normal. The bar offers nothing after High. Highest can only be set in the Dictionary Manager, because it also takes more swipes of similar words.

Only plausible corrections count (the chosen word must resemble the swiped one, so changing your mind about what to write is not recorded). The counts are part of on-device learning: they are kept only while Learn From My Typing is on, never in private/incognito or password fields, and are erased by Privacy & Data > Forget Learned. If the bar is busy when the second correction happens, the offer waits for the next idle moment.

When Swipe Typing Doesn’t Work

Swipe typing may not activate when:

• Swipe typing is disabled in settings
• Typing in password fields (unless enabled in settings)
• The swipe is too short (detected as tap)
• No language pack is available
• The swipe starts on a key that is not a letter. A word swipe must begin on a letter key. A swipe that starts on Backspace, Shift, Enter, Space, Ctrl/Fn, a digit or punctuation keeps that key’s own tap or short-swipe action, even if your finger then crosses letters (for example, a swipe-left from Backspace over m n b no longer types “mb”). A swipe that starts on a letter and later crosses the spacebar is unaffected.

In terminal apps the swiped word is inserted as recognised, without the final autocorrect step (so ls is not changed to is). See Advanced Settings.

Related Features

• Short Swipes - Quick access to subkeys
• Autocorrect - Fix mistakes automatically
• Per-Key Actions - Append ‘s / Append apostrophe and dynamic templates

Apostrophes and Short-Word Limits

A single apostrophe assigned as custom text follows ordinary punctuation behavior: automatic swipe spacing can be reclaimed, while manually typed spaces and spaces before a selected text range are preserved. A multi-character text macro such as 's remains literal. The separate Append ‘s and Append apostrophe commands attach to a verified word and support suffix-only undo; see Per-Key Actions.

Some short words, including reported ad and wet traces, remain recognition gaps in default CTC. Changing engines for each word is not the intended solution. Fresh human traces and general model calibration are needed before shipping a correction; passing automated routing tests does not establish short-word recognition accuracy. If you use one of these words, you can give it a swipe priority in the Dictionary Manager. Similar words then lose some swipes; for example, as loses some swipes to ad. Continuous multiword swipe is optional and disabled by default; enable it to use intentional spacebar dwells between words, as described below.

Continuous Multiword Swipe (development build)

Enable Continuous swipe under Settings → Swipe Typing. It is off by default. Swipe a word, hold in the center of the physical spacebar for about 280 ms, then continue to the next word without lifting. Each deliberate hold ends a word; lifting finishes the last segment. A quick crossing keeps the original single-word behavior. Subkeys, spacebar edges and empty space visits do not create words.

Segments decode in order through the current engine. The preceding word must be accepted before the next word is decoded. A deliberate spacebar boundary adds one separator even if automatic trailing spaces are disabled; the final word follows your normal spacing preference. Observed one-key English a and I segments are inserted directly; other segments require actual decoder candidates.

Changing field, layout, language or settings, starting another gesture, adding a second finger, or editing/moving the caret cancels pending words. Text already accepted remains. If a word was still being recognized — for example the last word, when you touch the keyboard or type a key right after lifting — the suggestion bar says the continuous swipe stopped, so check the text. Before your first spacebar hold, a cancellation simply leaves an ordinary single-word swipe. A rejected/empty decode or failed editor write stops the phrase with feedback; CleverKeys does not guess a replacement or retry an uncertain edit. Password fields, selected ranges, unreadable editors and active clipboard/emoji/GIF editors do not start continuous mode. With no intentional boundary, the normal single-word recognizer and its full original path still handle the gesture.

The new automated tests cover segmentation and queue behavior. Human phrase accuracy, spacebar hold comfort and latency still need device testing before release. These changes do not repair the separate shipped CTC ad/wet ranking problem.