What we believe about reading and listening

We never play an audiobook above 1×

A narrator’s timing is part of the performance. The pauses are not dead air the engineer forgot to cut. They are where a joke lands and where a revelation is allowed to arrive. Speeding a reading up is not a neutral act, and every runtime we quote is therefore the runtime as performed.

You are allowed to disagree with us about this, and plenty of people we like do. Tell us how you actually listen and we will grudgingly redo the maths your way.

We measure how you read, not what you like

Readers are not interchangeable. The same three hundred pages are four hours for one person and eleven for another, and every reading estimate on this site divides by a single middling pace because it has never had anything better to go on. So tell us roughly how fast you read and we will use your number instead of ours.

Here is the boundary, and it is a hard one. What we learn about you changes what the page shows you and never what the model thinks. The engine that decides read against listen is the same engine for everybody: it reads public evidence about books and no fact about the person looking at it. A reading pace moves one estimate on the page. It does not move a verdict, it does not move the slot we prescribe, and it could not: how long a book takes carries no weight in the call at all.

Your answer is kept in your own browser and goes nowhere near us. If you clear your history it is gone, and the page falls back to our 50 pages an hour with no harm done.

We will never sneer at audiobooks

For a great many people audio is access rather than preference: dyslexia, low vision, chronic pain, ADHD, ninety minutes of motorway each way. A book we tell you to read is not a book you are lazy for hearing, and the lazy version of this site’s voice is one that punches at exactly those readers.

We think 64 of our 198 works are better heard than read, and we mean it as praise. When a recording is the right answer we say so with actual enthusiasm.

We always show the working

Read it: the flashback structure needs a page you can flip back to lands. Trust us does not. Every verdict here is followed by the numbers it was read off, in the margin and in plain English, including the ones that undercut it.

That extends to admitting what the numbers rest on. Two of the quantities behind the call are directions we declared rather than measured, and we have never yet tested them against a real reader’s opinion. You will find us saying so on the pages themselves.

We are willing to shrug

Confidently wrong is worse than quiet. 22 of our 198 works come back Not enough to go on, and that is a real answer rather than a failure to produce one. Usually it means nobody has recorded the book, or so few people have rated the recording that any claim about it would be noise.

Either. Don’t overthink it is a real answer too. Some books genuinely do not care how you meet them, and pretending otherwise to seem useful would be the easiest way to become useless.

The verdict is about the book, not the recording

A different narrator does not change whether a book wants your eyes, so our call is always about the work. Where more than one complete recording exists and both are well attested, we name them and let you pick the voice.

Here is the whole of it as one rule. The stamp belongs to the book. Every judgement we make about a voice belongs to a recording we have named, and says two things: which comparison we made, and how many listeners are behind it.

There are exactly two comparisons and we never print a figure beside the wrong one. The first is how far a reading beats the book it is reading, set against every other book on our shelf: one reading a book, the one we picked for it, and that is the comparison our call actually reads. The second is how the reading stands against what the book’s own readers made of it, a different crowd rating a different thing, which we report as a cross-check and never score. Neither is a comparison of raw stars, and that matters more than it sounds: listeners mark a performance higher than a story, and Audible’s crowd marks higher than Goodreads’, so every figure is measured against the middle of its own crowd before anything is subtracted. It is why a reading can be the weak half of its book on a page where the two star figures underneath look level.

And where a different reading would genuinely change the answer, we say so in one sentence with both calls named, rather than letting one stamp stand for a choice you have not been told about. That is true of 2 of our 198 works today. It is rare because we hold a second complete reading of very few books, not because it is a small thing.

The exception cuts the other way. Naming a poor recording names a working narrator by their own name, so we only ever do that where the evidence is overwhelming, and we say how much of it there is.

Here is what overwhelming means, as a number: 250 ratings of the reading, counted across both Audible shops, British and American, because it is one recording and the two crowds rate the same performance. It is one bar doing four jobs: what a second recording has to clear before we will put it beside our pick, what a reading has to clear before we will call it the weaker half of a book, what it has to clear before we will draw its star charts or say how often the listen goes badly, and what it has to clear before a judgement about the reading is allowed to decide the verdict at all.

That last one is the newest and the bluntest. Under the bar we do not quietly discount the reading, we leave it out: the book is judged on its own pace, its own difficulty and its own print record, and the page says so and prints what the reading scored anyway. It cuts both ways, which is the point: a reading too thinly rated to be blamed is also too thinly rated to be credited. That is the case on 25 of our 198 works today.

The same bar decides what gets drawn. Where a recording has fewer ratings than that, its star charts fold away behind a line saying how many, because a bar chart of nine listeners is the same width on your screen as one of twelve thousand readers, and the picture makes a claim the caption underneath cannot take back. Open it and it is all still there. Above the bar, the performance chart also carries one figure the shop never prints: how many people gave the reading one or two stars, and what share of the room that is. It is there on every recording over the bar, well received or badly, because a number that only turned up when it was unflattering would be an accusation rather than a measurement. And it counts listeners, never the narrator.

What puts a recording in either hall

Two pages here are about a single recording rather than about a book. They are the same idea with opposite signs, and both of them start at the bar above: under it we say nothing about a voice at all, in either direction.

The Hall of Fame is a great book, superbly read. A recording gets in on the listening: a full step (one standard deviation, the typical distance a book sits from the middle of our own shelf) above that middle on Audible’s performance score and on its story score, with no more than 2% of listeners giving the performance one or two stars, on a book that is itself at or above the middle in print. Or it gets in on a prize: an Audie, won rather than shortlisted, in one of the categories that judge the narration itself, on a recording that existed in time to have won it. A genre prize like Fiction is given to a whole audiobook rather than to a voice, so it does not count. And even then, where our own listeners put the reading below the middle of the shelf, the prize does not carry it in alone.

The Hall of Shame is a good book whose recording is the weak half. It takes more to prove, not less, because it names a working narrator. Two independent measurements have to agree. Our own arithmetic must already call the reading the weak half of its book, which is the same judgement that tipped that book to read it in the first place; and an unusual share of the room must actually have bounced off it, more than the middle half of our shelf gets. The book still has to hold up in print, because the claim is a mismatch and not a bad book.

One rule sits across both and is deliberately lopsided. Any won Audie, in any category at all, keeps a recording out of the Hall of Shame outright, while only a prize for the narration itself is a way into the Hall of Fame, and even that can be overruled by our own listeners. A prize is weak evidence for praise and conclusive evidence against blame. We would rather be uneven about awards than wrong about a person.

Neither hall is ranked and neither is curated: both are listed alphabetically, both are computed from the same evidence as every verdict, and a recording can leave either one when the evidence moves. And a place in either belongs to a recording, never to a book, which is why we never put a series in one, and why a place says nothing about the verdict printed above it.

A synthetic reading is not a performance

A narrator’s timing is part of the book. We said as much at the top of this page, about the pauses a joke lands in. A generated voice has no timing to have. So we will never put one forward as the recording we mean: a machine-read edition is disqualified before any question of how good it is comes up, in the same breath as an abridgement. Where that leaves a book with no reading we are willing to name, we say we cannot call it rather than sending you to a machine.

What we are not saying is that a synthetic reading is worthless. If it is the only way you can meet a book, then it is the only way you can meet a book, and we would far rather you met it. We are telling you what the thing is, not what you should be able to do with it.

We do not read your ratings

Nothing you have rated, shelved or finished feeds the model. It is built from public evidence about books (how the crowds rated them, how the crowds rated the recordings, what prizes they won, how demanding readers found them) and never from a profile of you.

That is partly a matter of taste and partly a matter of proof: a model tuned to agree with somebody cannot then be tested against them. Ours has not been tested yet, and we would rather tell you that than imply otherwise.

What we do keep is the short list of things you have told us outright: how fast you read, whether you overrule us on playback speed, whether you want the ratings compared like for like, day or night. They live in your browser, they change what a page shows you, and not one of them reaches the model. That line is the whole of our position on personalisation: the rendering can be yours; the opinion is ours, and it is the same opinion we give everybody.