Keyboard Swipe Distance

iOS 13 introduced a swipe keyboard (officially called QuickPath) to the base operating system. This has been available in other operating systems as well as in various third-party keyboards for some time. However, I’m unadventurous and so this was the first time that I’ve tried it out. I’ve been enjoying it but found that some words seem far less efficient to type by swiping (for whatever that means on a tiny, mobile keyboard).

I’ve written a Go package and small CLI to calculate the swipe distance required to type words. It does some simple maths based on the pixel coordinates of key in a screenshot of a keyboard. Using this program we can calculate the total swipe distance required to type all of the words in the dictionary. A caveat is that there may be shorter ways to type some words if they have a unique swipe prefix so that the word can be identified without completing the swipe. This will be ignored for these calculations.

Let’s start by just getting the top 10 longest-to-swipe words from the default dictionary on my machine.

$ cat /usr/share/dict/words |
    swipedistance |
    sort -k 1 -n |
    tail -10
7062.577492	307.068587	pseudolamellibranchiate
7066.836197	336.516009	pentamethylenediamine
7092.310552	354.615528	paraphenylenediamine
7100.013821	338.095896	pseudohermaphroditism
7112.534633	323.297029	electroencephalography
7159.586349	311.286363	phenolsulphonephthalein
7201.797546	313.121632	Pseudolamellibranchiata
7239.048156	314.741224	scientificogeographical
7246.974277	301.957262	formaldehydesulphoxylate
8162.285953	388.680283	mandibulosuspensorial

Those… are some impressive words. Unsurprisingly long words require more swipe distance in order to type them. This isn’t exactly a revelation. What may be more interesting is words which require a disproportionate amount of swiping for their length. The second column in the output divides the total distance by the number characters in the word.

$ cat /usr/share/dict/words |
    tr 'A-Z' 'a-z' |
    uniq |
    swipedistance |
    sort -k 2 -n |
    tail -10
4455.781192	495.086799	malayalam
3502.734049	500.390578	amalaka
3002.451554	500.408592	palame
2017.697946	504.424486	papa
4614.467659	512.718629	anapanapa
3081.300031	513.550005	palama
2588.000000	517.600000	alala
3109.131964	518.188661	jalapa
2664.697946	532.939589	papal
5124.844360	569.427151	palapalai

Note: there were some duplicates in the output without the additional tr and uniq parts of the pipe.

This list is a little more interesting! We can see words like “papal” near the top of the list which makes a lot of sense when we look at the keyboard. The word requires four full-width swipes across the keyboard in order to type it. All of the other words in the list seem to follow a similar pattern: they all contain long substrings of alternating letters on the extreme left and right of the keyboard. Let’s do another check but exclude characters on the far left and right of the keyboard.

$ cat /usr/share/dict/words |
    tr 'A-Z' 'a-z' |
    uniq |
    grep '^[^qazplm]*$' |
    swipedistance |
    sort -k 2 -n |
    tail -10
3365.101010	373.900112	toxicosis
3418.436110	379.826234	crownwork
3048.202226	381.025278	widowish
2675.409214	382.201316	toxosis
1544.559598	386.139899	sown
2706.518534	386.645505	sorosis
3096.161916	387.020240	downtown
1162.358662	387.452887	lwo
1557.439244	389.359811	soso
2354.619313	392.436552	disown

These words are a little more varied and starting to get pretty interesting! We do see common patterns again e.g. “so” and “is”. There’s probably some filtering we could do in order to filter out these substrings but I’m worried that we’d eventually just regress to the longest word wins situation in our first example.