words and n-grams

An n-gram is n letters in a row inside a word: th is a bigram, the a trigram, tion a tetragram.

Skilled typists control whole words: words in random order are typed almost as fast as real text, within about 3% (Salthouse, 1986), while scrambled letters slow them down sharply (Yamaguchi and Logan, 2014). Practice gains stay close to what was practised: after training whole words, new words made of the same letters got only about 4% faster, and that faded within the session (Crump and Logan, 2010).

So from about 40 wpm most practice belongs on real words, with n-grams as a short block for your weakest letter steps. That is our advice and a working hypothesis: no study has compared the two for experienced typists. Below 40 wpm the frequent bigrams and trigrams are a good way to learn the common motions.

A few items do most of the work. The share of English the most frequent ones make up:

Share of all letter pairs, triples and runs of four inside words in Google Books, and of all 51 million words said in film and TV subtitles. th alone is 3.6% of all letter pairs. Auto scope starts with the top 20 bigrams, about a third of all pairs. The steps into and out of words, which these counts leave out, are about a third of all steps between two keys.

the data

Two real sources, used as they are: every item carries its real count, nothing is generated.

n-grams

In 1965 Mark Mayzner published letter-combination frequencies from a sample of 20,000 words. Peter Norvig repeated the count on Google Books: 97,565 distinct words, mentioned 743 billion times. Letters a to z, counted inside words

words

SUBTLEX-US (Brysbaert and New, 2009): how often words are said in American film and TV subtitles, 51 million words. Its counts predict how fast people recognise words better than counts from books do (Brysbaert and colleagues, 2011). Spoken words run shorter than written business English, so your own texts (custom) fit your writing better. Names (found by part of speech), sounds such as mmm, pieces of contractions, two-letter debris and strong profanity are left out; days of the week and months stay

written

written English words by how often they appear in books: Peter Norvig's counts from Google Books, 97,565 words, the top 10,000 kept. Words run longer than in subtitles and closer to business writing. The list has no capitals, so names are found from case in real sentences and dropped; a few may remain far down the list. Loaded the first time you choose it

sentences

real English sentences with capitals, commas, full stops and other marks, now and then digits: Tatoeba, the sentences released under CC0, 40 to 100 characters, every word among the 10,000 most frequent written words. A collection of example sentences, not a frequency sample, and more conversational than business text. A fifth of them are kept aside for the capitals check and never come up in practice. Loaded the first time you choose it

your texts

in custom, paste a document of your own and press count my text: its words are counted in this browser, the text itself is not kept, and you review the list before saving it. Only the words with their counts are saved, and they sync with your progress, so delete client and people names first. A counted list is practised like words: auto scope from your 50 most frequent (set in the list panel)

counts

also weigh the drill: a frequent item matters more than a rare one

Sources: Google Books Ngram data, CC BY 3.0, as tabulated at norvig.com/mayzner; Tatoeba sentences released under CC0; SUBTLEX-US by Marc Brysbaert and Boris New, CC BY-NC-SA 4.0 (the licence file of the OSF project). The word list in thhe is a filtered version of SUBTLEX-US (filters above), shared under the same licence, for non-commercial use.

how it adapts

measure

every two keys in a row, thhe records the time between them. Word starts (the step from the space to the first letter) take longest, and that time belongs to starting the word, so n-gram items are timed and graded only on the steps inside them, and word starts are measured on their own; words are timed from the space before them

clean

only first-try keystrokes count for speed; a miss belongs to the step that leads to the missed key; when one left-out, extra or swapped key shifts the keys after it, the slip counts once and the shifted keys are not charged; a key in the other keyboard layout counts as a wrong key but charges no letter step; a hesitation counts up to 2 s, a gap over 5 s is a pause and is not timed

recent

speed follows your recent tries; misses are judged on the long run, so one bad phrase does not label an item

once

an item repeated inside one phrase counts as one try, so mastery needs several phrases

borrow

an item you have not typed yet takes its estimate from the letter steps inside it, so a word full of your slow steps comes up before you meet it

weigh

slower than expected at your usual pace and more so above the bar, misses above your own average, more frequent, still new: each raises the weight; in words a slow word start adds a little

pick

about 70% of a phrase is drawn by weight, the rest at random from the scope, so strong items stay fresh. A fixed word scope above 200 draws the weighted part from a sliding working set: the first 200 words of the scope that are not mastered yet, so records build up for each word

words

for every word you type thhe keeps its own record and picks by it within the scope you choose, 50 to 10,000 words; your own list works the same way. The record is one per word, shared by words, written and custom: a word you miss in one list comes up as weak in the others too, and sentences come up more often when they hold your weak words. This is the owner's design: other trainers adapt by letter, by weak letter steps inside words, or replay the words of the last test

master

met in 4 timed phrases or more, recently faster than the bar, and a long-run miss share not clearly above your own average for the source (before a source has about 100 tries: within your accuracy target). The stats table says what holds an item back: tries, speed or misses

unlock

with scope on auto, the next most frequent items join once 80% of those in play are mastered; new ones are measured first, so they come up often

recheck

mastery on one day can be that day's form, so the stats mark it unconfirmed until it holds again on a later day; mastered items come back now and then at growing day intervals (1, 2, 4, 8 days and so on)

confuse

thhe notes which key you strike instead of the right one. When one wrong key clearly repeats for a letter, words phrases now and then include a frequent word with both letters, until the confusion is no longer above your usual rate. Whether this helps has not been tested

how to train

1. Speed and accuracy. The only research comparison for typing is old: in typewriter-era classes, long periods of speed practice followed by shorter accuracy work did better than heavy accuracy practice, and strict early accuracy standards were among the useless procedures (West, 1962). Practitioners mostly say accuracy first. No modern trial with experienced typists exists. Our advice, as a trial: practise mostly a little above your usual pace (wpm auto) and accept misses there; add short clean work at a relaxed pace; fix a felt slip at once. Every wrong key counts, fixed or not.
2. Target your pace. With auto on, the wpm target follows your recent pace, the median of your last 50 phrases plus 2 wpm or 5%, set once a day; the bar for mastery rises to it only when 80% of the items in play are mastered. Turn auto off to set your own.
3. Words from about 40 wpm. Below that, learn the frequent motions on bigrams and trigrams. From about 40 wpm, spend most of each session on words with adaptive on, and a short block on your weakest steps: bigram or trigram phrases, or words with adaptive: pairs. Leave scope on auto, or fix a word scope; above 200 it practises a sliding working set (our advice).
4. Read the trend, not the phrase. Single phrases vary by several wpm. Once a week thhe offers a check: about five minutes of the 1,000 most frequent words, the same kind every week, with a warm-up, no clock and no retries. Its trend over weeks is your progress number; from the 9th check stats says rising or falling only beyond the change your own checks can detect. Between checks, the median and its ± after each phrase, the speed line and the day-start line show change; drill pace on n-grams runs faster than real text.
5. Your own words. General word lists miss much of a working vocabulary. Count a few of your own documents into custom and practise them like words (our advice; practice speeds up the words practised, Crump and Logan, 2010).
6. Drill, then flow. A few repeats inside a phrase build a motion; mixing items across phrases makes it stick. See the recommended settings below.
7. Eyes on the text, not on the keys (our advice; attention to the hands slows skilled typing, Logan and Crump, 2009).

recommended settings

Repeating one item many times in a row gives fast gains during practice, but mixed practice was retained better in lab motor tasks (Shea and Morgan, 1979); in applied settings the benefit is smaller (Brady, 2004), and no study has compared the two for typing. So each phrase repeats a few items a little, and the phrases themselves mix.

bigrams

length 3, rep 3: about 25 characters

trigrams

length 3, rep 2

tetragrams

length 3, rep 2

words

length 25, each word once: a long stretch of real words, close to real text; written the same

later

when a source feels easy, try length 5 to 6 and rep 2 for more mixing

these are the start values; once you change length, rep, marks or a target, each source keeps your own value across switches and reloads.

how long

daily

short daily sessions rather than long ones: postal workers learning a keyboard learned better with one hour a day than with two sessions of two hours a day (Baddeley and Longman, 1978). Nothing shorter than an hour was tested, and no typing study has compared 15 and 60 minutes a day, so the 10 to 20 minutes are our suggestion, not a research figure

goal

set a daily goal in minutes (15 by default): the line under today's minutes at the top fills as you type; only active typing counts, pauses of 5 s or more do not; the day starts at midnight; stats shows your streak and the last 30 days

stop

when your hands tense up (our advice). A slowdown late in a session is a passing effect and does not harm learning

expect

gains on real text are slow and uneven, roughly 0.3 to 1 wpm a week at 15 to 20 minutes a day, possibly less if drills carry over poorly (an estimate from practitioners' logs, not a study). Call a plateau only after a flat trend over two or three months

check

the weekly check takes about five minutes; check keys never enter your practice statistics. Stats shows each check, a trend in wpm a week from the 4th, and your slowest letter steps in practice against the same steps in the checks. Once a plain check exists, stats also offers a capitals check on sentences kept aside from practice: it measures what capitals, marks and digits cost you, step by step, so you can judge whether practice on sentences is worth your time

writing

in long documents most time goes to thinking and sources, not keys: 5 to 10 wpm more saves at most about 1 to 7% of writing time, and fewer misses may buy about as much (an estimate)

warm-up

letter-step timings fill up fast, every keystroke counts; an item needs to come up in 3 phrases before it stops being new; larger scopes take several sessions, and until then their items show as new and come up more often

the interface

type, listen

type: the text is on screen, as in every source below. listen: dictation; a recorded voice (Kokoro, open source) says short sentences of the 1,000 most frequent words one at a time, at its natural speed or slower, with pauses that keep your target wpm; the text stays hidden and you type what you hear; as a step towards that, text after shows each sentence once it has been said, and text live shows the words as they are said. Capitals and punctuation are not checked. Its errors stay out of the typing statistics, its active time counts for the daily goal, and stats shows your last runs. It is the closest practice to taking notes in a meeting; that it carries over is our reasoning, not a measured result

marks

for words, written and custom: off, real (punctuation and capitals at the rates of general written English) or more (about twice as many, for drilling); set basic is the full stop, comma, question mark and possessive apostrophe, full adds the rarer marks. Your word records stay letters-only, so a capital or a comma never makes a word look slow; the steps into capitals and marks show in stats by kind

source

bigrams, trigrams, tetragrams; words (spoken) or written (books); sentences with capitals and marks, one a phrase, recorded only as steps between keys; or custom: your own lists, or the words of your own texts with their counts

custom lists

up to 30 named lists, synced with your progress; the list name in the bar opens them, to switch, make, rename, edit or delete one. Words go split by spaces or new lines; several lines with a space in one of them are phrases, one per line, each kept and timed as a whole (up to 40 characters). Curly quotes and dashes become the plain keyboard ones. Order, per list: weak (your slow and missed items more often), shuffle (each item once a round) or in order (from the top, carrying on where you stopped, round after round; a phrase moves the place on once it is finished)

scope

how many of the most frequent items are in play; auto starts with the top 20 (words: 50) and adds 10 more (words: 50) each time 80% of those in play are mastered; a number fixes the scope, and a word scope above 200 practises a sliding working set of 200 words that are not mastered yet

length

items per phrase; each source keeps its own

rep

how many times the phrase repeats them; each source keeps its own. Not offered where phrases run once: words, written, and a custom list in the weak order, where weak items come back in the next phrases instead

wpm, acc

the target: the red mark and the bar for slow, weak and mastered items. wpm auto follows your recent pace; acc is judged over your last 200 or so keystrokes (at 100, phrase by phrase)

adaptive

on: picked by weakness; off: the whole scope in random order, each item once per round; on words a third state, pairs, drills your weakest letter steps inside real words

typing

grey letters fill in as you type; red is a miss; fix it with backspace (alt + backspace deletes a word); the phrase ends on a right last letter, so a wrong last key waits for a fix; for 0.3 s after a phrase a repeat of its last letter is ignored, so a key pressed by momentum does not land in the new phrase; a key in the other keyboard layout is not typed; misses left unfixed lower the accuracy; a held key types once; 10 s without a key ends the attempt and a new phrase follows (what you typed still counts); on a phone tap the text to open the keyboard; phone typing stays out of the keyboard statistics

below

wpm (drill pace on n-grams) and acc of the last phrase with its misses, the accuracy over your last keystrokes, and the median of the last 20 phrases with its noise; focus: the items or letter steps the lesson leans on now; seen: items typed at least once; mastered: progress to the next auto step; starts: your slowest word starts

time lost

under the goal of the day, where your last phrase's time went (before the first phrase of a visit: the last 7 days): for every step between two keys, the time beyond your usual time for that step, in four kinds: fixes (steps with a backspace: the wrong keys, the backspaces and the retyping), stops after misses (a long step right after a wrong key you left, or a stray Enter), pauses between words and pauses in words (2.5 times the usual step or more), with the words where they happened; the costliest kind is yellow. The kinds show when they add up to 5% of the phrase or more, else the total in grey; a slip you make often shows as e → r again

tips

when one kind of lost time leads over your last 10 phrases of a source and takes 4% of your typing time or more, a tip under the result names its lever: fewer slips rather than faster fixes; fix at once or type on; read one word ahead; or the hard letter steps the model drills. At most once every 5 phrases

goal of the day

the first line under the phrase: one goal a day for the source (and one for listen), from your 7 days before today, so it holds all day: accuracy below your target first, then the biggest kind of lost time, then slow word starts, and with all of it clean (accuracy a point or more above your target) a push for speed, else steady; today's number in brackets, met after 10 phrases at or past the aim. Hover any of these lines for how it is worked out

keys

tab restart (not in the weekly check) shift+enter new phrase, or stop the check esc stats and back shift+tab move to the controls; on a hint, enter opens it

stats

weakest first; grey speed is an estimate not yet measured, red is slower than the bar; misses is the recent miss rate per character, red when clearly above your average; the grey words say what holds an item back, and unconfirmed marks mastery seen on one day only; the bar shows how strongly an item is picked. Above the table: the weekly checks, one row per day of the source (word start and inside-word step times, misses, characters left wrong, corrections and their time; older days add up into weeks), where the time goes, slips by type, and your most frequent confusions

progress

on thhe.app: sign in with Google and progress follows you to every device. It syncs every 30 s while you type and when you leave the page; anything not yet synced waits in this browser and is added later, and two devices merge rather than overwrite each other. After an update, a page opened before it asks for a reload before it saves. Progress made signed out joins your account at sign-in. Stats has sign out everywhere and delete account. On this Mac with thhe.command: progress.json next to the app. Export and import move progress by hand