Towards better typing apps

I’ve been experimenting with a better way of generating random text for typing tests. You can try the results of the experiment here, or read my ramblings below.

A frustration with existing typing apps

Many typing practice apps provide the option to type a random sequence of words, whereby words are drawn uniformly from some small set. Commonly the most frequently used 200 words are used. The larger the source set, the harder the test is, because we’re forced to type more obscure and less well practiced words.

Zipping effortlessly through the most common 200 words is satisfying, but isn’t great practice for typing real text. On the other hand typing tests of this sort that draw from a large enough set of words feel wrong. They become harder work than typing real text, because seemingly every word is long and obscure, and the usual rhythm of typing mostly common words punctuated by the occasional zinger is lost. Real text contains a huge proportion of a small number of very common words (see Zipf’s law).

A better way

I propose a better solution: generate text with realistic frequencies. If “the” accounts for 7% of all words in the wild, it should also make up 7% of the words in your typing test. The results feel right and it’s also perfect practice, in so much as how often you practice a word will be exactly proportional to how often it appears in actual use.

This is technically easy to achieve given a frequency list such as this one on Wiktionary, which lists words from a large corpus along with their frequency of occurance. Instead of selecting words uniformly at random from the full set, take a weighted selection using the frequencies as weights. I’m using the top 10,000, which accounts for 97% of all words.

To get a feel for this, here’s a uniform selection from the top 200 words:

would be there day yes you’re like still what it don’t your long do those how mean like you’ve let’s with day had why isn’t well I’d let anything into what one and these about great him was like them

1000 words:

face somebody book mouth huh clothes isn’t seen alive so Todd eight hey Christmas present also police self person works hang president kill before ready clothes hold we she’s yeah part lately day big

10,000 words:

council bubble manticore given invitations recently episodes moms valentine Maxine anything’s leg fears gettin unbelievably lip dancer bigger onion supposed taking hook invitations Madeline sinking trying

and finally a weighted selection from the top 10,000 words:

high a you’re it the your and the make while me so the can’t maybe our because a he me kill sister with went any has well do know for for mean stand of hooker a is actually just right used means you of

Under the hood (for those who care about such things)

The whole app is about 5kb of Javascript – written by hand with no compilation steps and no dependencies. It’s a fun way to work sometimes. The ui is a “React-ish” function of the app state, but I just throw away the whole dom and build it from scratch on every key press. It’s plenty fast enough because there’s really not much going on, but maybe I’ll write a vdom implementation one day…