Replicated across many independent laboratories over decades. Safe to build on.
Reading has been studied intensively for over a century, and most of what the speed-reading industry sells is contradicted by it. This page separates the mechanisms we can defend from the ones we're still testing, and states plainly where the evidence runs against us.
Replicated across many independent laboratories over decades. Safe to build on.
Robust effects in the literature, applied here in a way that goes slightly beyond what was directly tested.
Our own reasoning. Consistent with the research, but nobody has tested this specific combination — including us.
A competent adult reads continuous prose at roughly 200–300 words per minute. That number has been remarkably stable across a century of measurement, which should tell you something: it isn't a bad habit waiting to be broken. It's what the machinery costs.
Reading is not a smooth glide along the line. The eye fixates for roughly 200–250 milliseconds, jumps forward seven to nine characters, and fixates again. During the jump you're effectively blind. Around ten to fifteen per cent of reading time is spent in motion, taking in nothing.
On top of that, skilled readers move their eyes backwards on roughly one saccade in seven. Some are corrections for overshooting. Many are comprehension failures: the sentence didn't resolve, so you go back and try again.
The word comes to a fixed point instead of the eye going to the word. No saccades, no return sweep, no place in the paragraph to keep. That reclaims the motion cost outright — the one saving in speed reading that's genuinely mechanical rather than wishful.
Most readers silently articulate what they read. Subvocalisation is measurable as real activity in the speech muscles, and it isn't a defect — it supports comprehension, particularly for difficult material. But it chains reading rate to something close to speaking rate, which tops out far below what the visual system could deliver.
The classic advice is to suppress it. The honest finding is that suppressing it works, and it costs you comprehension on anything demanding, because the phonological route was doing useful work.
We don't ask you to suppress anything. Above roughly 400 wpm subvocalisation falls away on its own because there's no time for it — and the paired image is there specifically to carry the load the inner voice was carrying. That substitution is what the whole app rests on.
This is the finding that ends most speed-reading claims. When researchers remove eye movements entirely and present words one at a time at a fixed point, reading gets somewhat faster — but nowhere near the promised multiples. The limiting factor isn't how fast you can see words. It's how fast you can integrate them into meaning.
Any product claiming a fourfold gain with comprehension intact is either measuring skimming or measuring nothing. Removing eye movement buys real but modest time; after that you're arguing with the language system itself.
We accept the bottleneck and attack it from a different side — rather than pushing symbols through the language channel faster, we send meaning down a second channel that isn't queueing for it. Whether that fully works is the open question of this entire project. See what we will not claim.
The strongest evidence behind this app has nothing to do with speed. It concerns memory, and it's one of the most durable findings in cognitive psychology: information encoded both verbally and visually is remembered substantially better than either alone.
Verbal and imagery-based representations are partly separate systems that can refer to each other. A concrete word like river activates both; an abstract word like criterion mostly activates one. That's precisely why concrete words are consistently easier to recall — they arrive with two retrieval routes instead of one.
Pairing a word with a meaningful image builds the second trace deliberately, for every word, including the abstract ones that would never have got one.
Every word in the stream — not the interesting ones, not the nouns, every word — is paired with an image chosen by sense. Abstract vocabulary gets a picture through morphology and composition, which is exactly where dual coding normally has nothing to offer.
Show someone a list of words and a set of pictures, then test recognition later. The pictures win, and by a wide margin. Human memory for images is extraordinarily capacious — people shown thousands of pictures in a single session recognise them at rates close to ceiling days afterwards.
Nothing comparable is true of text. There's no experiment in which people recognise ten thousand sentences at ninety per cent accuracy.
Every image is a distinct, hand-drawn or procedurally generated mark rather than a generic pictogram, so words stay visually discriminable. A procedural glyph is derived from the word itself — even an invented word draws the same distinctive shape on every machine, every time.
Understanding action verbs recruits parts of the brain involved in producing and perceiving those actions. Motion isn't decoration attached to a static concept; for a large class of words it's part of the concept. Drip, flicker, surge and collapse are defined by how they move.
Animation also has a cost: motion captures attention automatically, and unnecessary motion competes with the thing you were trying to read.
Twenty-eight parametric motion primitives, assigned by meaning and able to target sub-elements so only the raindrops fall. Amplitude is a single control that scales every animation at once, down to zero — and prefers-reduced-motion silences all of it unasked.
Once a word is alone at a fixed point, two decades of eye-movement research still apply to it — because where the eye lands within a word, and how long it needs to stay there, are properties of the word rather than the page.
Word recognition is fastest and most accurate when the eye lands slightly left of the word's centre — roughly a third of the way in. Land elsewhere and recognition time rises measurably. In natural reading the eye targets this position without being told to.
In a fixed-point display this matters more, not less: if words are simply centred, the recognition point shifts with every word length and your eye makes tiny corrective movements you never asked for.
Each word is offset horizontally so its optimal recognition point sits exactly on the centre rail, and that character is marked in amber. Across the whole stream your eye holds one position with no correction at all.
Fixation duration isn't constant. Long words take longer. Rare words take longer than common ones. Hard-to-predict words take longer than predictable ones — and short function words are frequently skipped altogether. Punctuation and clause boundaries add integration time, because that's where a sentence is assembled into a proposition.
A constant-interval word stream ignores all of this. It gives the and notwithstanding the same 150 milliseconds, which is generous to one and cruel to the other.
Dwell is modulated per word: up to 1.5× for long words, 1.35× for numerals, 1.6× at commas, 2.2× at sentence ends, 2.8× at paragraph breaks. Your true rate therefore runs slightly under the nominal figure on dense prose. We'd rather report that than hide it.
If the image is wrong, the second channel is worse than absent — it actively competes with the sentence you're reading. Everything else depends on getting the sense right, which is why it's the part we spent the most time on.
Ambiguous words briefly activate more than one meaning before context selects the winner. Where two meanings are roughly equally common, resolution is measurably slower — a real cost, visible in fixation times. Where a prior context strongly favours one sense, that cost largely disappears.
Which means a correct image supplied before or alongside the word isn't decoration. It's doing the disambiguation work the sentence would otherwise charge you for.
Every sense carries context cue words, scored against an eight-word window either side of the target, with frequency rank breaking ties. 125 of the most treacherous words in English are disambiguated this way — bank, bat, spring, light, present, fine and the rest of the usual offenders.
A word is recognised faster when a semantically related concept has just been activated. Prime with doctor and nurse arrives quicker. The effect is fast, largely automatic, and one of the most reproducible results in the field.
An image is an unusually direct prime, because it activates the concept without passing through the word form at all.
The icon flashes, then word mode shows the image alone for the first 35% of each frame, then reveals the word. It's a deliberate test of semantic pre-activation, and it's the mode most often reported back to us as the one that suddenly clicked.
English is largely head-final in its compounds: a riverbank is a kind of bank, sunlight is a kind of light. Morphology is similarly systematic — -less negates, -er makes an agent, re- repeats. These regularities are well documented in linguistics.
Whether a composed image is understood as fluently as a drawn one is, as far as we know, not something anyone has measured. We think it is; we can't prove it.
Tiers 2 and 3 of the resolver compose images from parts and overlay morphological modifiers, and the app labels every image with the tier that produced it. If composed images turn out to be weaker, you'll be able to see exactly which ones they were.
Working memory isn't one pool. There's a limited store for verbal-phonological material and a separate one for visual-spatial material, and loading both isn't the same as loading one twice as hard. It's why a diagram narrated aloud is learned better than the same diagram with the same words printed beside it.
The image occupies the visual channel while the voice occupies the auditory one, both locked to a single clock. Speech runs a sentence at a time with the audio as master timer, so drift can't accumulate, and the next sentence is synthesised while the current one plays.
The finding above has a well-known partner, and it cuts the other way. Presenting the same verbal content simultaneously as printed text and as speech can hurt rather than help, because the two must be reconciled while competing for the same verbal resources.
So we won't pretend turning on the voice is free. For some readers and some material it's a clear gain; for others it's a distraction. Anyone telling you otherwise hasn't read the same literature.
The voice is off by default, and one key away. There are two sync modes — reader-led, where your pace governs the voice, and speech-led, where the voice governs the display — because which one helps depends on why you turned it on.
Dyslexia is a difficulty with decoding rather than with intelligence or comprehension of spoken language. The research base is large, active and genuinely contested in places — so this section is deliberately conservative about what follows from it.
A letter is harder to identify when flanked closely by other letters, and this crowding effect is measurably larger in many dyslexic readers. It's a robust perceptual finding with a practical corollary that has been tested directly: increasing letter and word spacing improves reading speed and accuracy in dyslexic children without any training at all.
Single-word presentation is the limiting case of increased spacing. There are no flankers, because there's nothing else on the screen.
One word, alone, at a fixed point, with empty field on every side. Not an accessibility mode bolted on afterwards — it's simply how the app works.
Struggling readers make more fixations, more regressions and more line-tracking errors than fluent readers. Some of that is a consequence of decoding difficulty rather than a cause — but effort spent keeping your place is effort not spent on meaning, whichever direction the causality runs.
Remove the line and you remove the class of error entirely. You can't skip a line that doesn't exist.
No lines, no return sweeps, no page. Pace is fully controllable in both directions, and plenty of readers run this at 120–180 wpm — slower than ordinary reading — purely for the fixed point.
If the bottleneck is turning letters into sound and sound into meaning, then a channel that delivers meaning without requiring the first two steps ought to help. That's the reasoning, and it's consistent with everything above.
It has not been tested on dyslexic readers, by us or by anyone else, and we will not describe it as a treatment, an intervention or a therapy. It's a way of presenting text. Some readers find it easier. Some don't. It's free, it needs no account, and nothing about your reading leaves your machine — so the cost of finding out is an afternoon.
icon replaces word turns content words into pure image while function words stay as text. icon flashes, then word primes the sense before the spelling arrives. Both exist specifically so meaning can lead and decoding can follow.
And they only mean anything if the image is actually there. A route around decoding that opens for common words and closes for rare ones returns the reader to the spelling at the exact moment it costs most, so the resolver was built to reach hundreds of thousands of words rather than a comfortable core — five tiers, tried in order, ending in one that cannot fail. That is an engineering guarantee about coverage. It is not evidence that the route helps, which is what the label above this card says.
Being tested on material produces better long-term retention than spending the same time studying it again. The effect is large, extensively replicated, and holds even when learners themselves predict the opposite — re-reading feels more productive, which is exactly why it survives.
So the comprehension check at the end of a passage isn't overhead. It's one of the most effective things in the entire app, and it doubles as the honesty mechanism.
Every passage ends with a recall check mixing words that genuinely appeared with plausible distractors drawn from the lexicon. The score is stored beside the speed, so you can see where increasing pace began to cost you and settle where it doesn't.
Readers routinely overestimate how much they've understood, and the illusion is strongest when material is fluent and easy to process. Speed reading is an almost perfect machine for producing that illusion: the experience is smooth, the pages advance, the feeling of understanding is intact — and it isn't necessarily connected to anything.
Words read counts the furthest point reached, so rewinding never inflates it. Session export writes a plain-text report with your rate, settings, resolver breakdown and comprehension history. It's deliberately unflattering, because a flattering number would be worthless.
Comprehension doesn't hold flat as rate increases. It declines, and past a certain point it collapses. Any honest account of speed reading has to start by drawing this, because everything useful happens in the region before the fall.
Illustrative, not measured. The solid line is the shape the literature consistently reports: comprehension holding near ceiling through the normal range, sagging through the middle, falling away entirely once you're sampling rather than reading. The dotted line is the shape we're trying to produce with a second channel. We have not demonstrated that we produce it. It's drawn as an aspiration, clearly marked, because drawing it as a result would be exactly the dishonesty this page exists to avoid.
A page about evidence is only worth reading if it's willing to list the things it can't support. Here's ours.
You won't, and nor will you with any other tool. The language system is the bottleneck and it doesn't have a setting. Expect a real but moderate gain — larger on familiar material, smaller on dense material.
Dual coding is well supported. Word-by-word presentation is well studied. Combining them, at speed, with automatically resolved images has not been tested in a controlled trial. That combination is our bet, not a finding.
It doesn't, and we'll never say it does. Several mechanisms point the right way and some readers report it helps them. That's a reason to try it, not a clinical claim, and no substitute for proper assessment and support.
Going back is sometimes how comprehension is repaired. Take it away and you lose the repair mechanism along with the wasted time. That's why there are step-back keys and a restart, and why we suggest a slow second pass on anything that matters.
Sense disambiguation covers 125 words deeply and everything else by frequency. Tier 4 pulls from open sets of varying quality, and tier 5 is a procedural mark with no semantic content whatsoever. Every image is labelled with its tier precisely so you can catch us.
Poetry, proofs, contracts and anything you intend to argue with deserve slow, recursive, sceptical reading with a pen. This is a tool for volume and triage. It isn't a shortcut around thinking, and we'd rather say so.
Every claim on this page is about people in general. The only question that matters to you is what happens with your eyes, your material and your memory — and you can answer it in about forty minutes, for free, without telling us anything.
Same source, similar difficulty, roughly a thousand words each. Something you haven't read before, and representative of what you actually read — not a children's story chosen to flatter the result.
Run the first passage in plain word-only mode at a pace that feels normal. Take the comprehension check immediately afterwards and write down both numbers. This is the only number you'll be comparing against.
Run the second passage with images on, at the same pace as your baseline. Same pace — this measures whether comprehension improves, not whether speed does. Take the check again.
New passage, images on, pace raised by 100 wpm. Check again. Repeat, adding 100 each time, until the comprehension score falls below your baseline. The pace just before that is your working rate.
The session report contains every run, every setting and every score. If the second channel did nothing for you, it'll say so plainly — and we'd still rather you had the truth than the subscription.
Two cautions on your own results. First, the second passage benefits from having warmed up on the first, so alternate the order if you run this more than once. Second, a single trial on a single afternoon is a data point, not a finding — the same caution we apply to ourselves throughout this page.
These are the literatures the page draws on, with a well-known entry point into each. They support the mechanisms described above. None is a study of this app, and none should be read as an endorsement of it.
Covers fixations, saccades, regressions, subvocalisation and word-by-word presentation, and explains carefully why the large advertised gains aren't attainable.
Rayner, Schotter, Masson, Potter & Treiman (2016), “So Much to Read, So Little Time”, Psychological Science in the Public Interest
Fixation durations, saccade lengths, the preferred viewing position, and the effects of word length, frequency and predictability on how long the eye stays.
Rayner (1998), “Eye Movements in Reading and Information Processing: 20 Years of Research”, Psychological Bulletin
The theoretical basis for pairing words with images, and the account of why concrete words are recalled better than abstract ones.
Paivio, dual coding theory — see “Mental Representations: A Dual Coding Approach” (1986)
The classic demonstration that recognition memory for large numbers of images is close to unbounded, in a way that has no verbal equivalent.
Standing (1973), “Learning 10,000 pictures”, Quarterly Journal of Experimental Psychology
The modality and redundancy principles — why narrated visuals beat printed-plus-spoken duplicates, and where adding a channel starts to cost you.
Mayer, “Multimedia Learning”; Sweller and colleagues on cognitive load
The evidence that being tested beats re-studying for long-term retention, including the finding that learners consistently predict the reverse.
Roediger & Karpicke (2006), “Test-Enhanced Learning”, Psychological Science
A direct demonstration that increasing letter spacing improves reading speed and accuracy in dyslexic children — the crowding result that single-word presentation takes to its limit.
Zorzi et al. (2012), “Extra-large letter spacing improves reading in dyslexia”, PNAS
Multiple senses activate briefly before context selects; balanced ambiguities cost measurable time, and a biasing context removes most of that cost.
Duffy, Morris & Rayner (1988) and the subsequent subordinate-bias literature
We've summarised these from the general literature rather than re-running the studies, and any error of emphasis is ours. If you find a place where we've overstated what a result supports, tell us and we'll change the page — considerably cheaper than being wrong in public.
Forty minutes, two passages and an honest score will tell you more about whether this works for you than any page of citations ever could — including this one.