Synthetic narration has got good enough that the interesting question is no longer “does it sound like a robot” but “can I listen to it for eleven hours”. Those are different bars.
What it does well
Sentence rhythm, mostly. A decent model puts the stress in the right place, pauses correctly at a comma against a full stop, and does not run questions into statements. Over a paragraph of ordinary prose it is genuinely unremarkable, which is the highest compliment available.
Non-fiction comes out well. History, science writing, essays, biography, technical books — anything where the job is to convey sentences rather than perform them.
What it does badly
Character voices. There are none. Every speaker sounds the same, so a page of unattributed dialogue between two people becomes hard to follow in a way it never is on paper.
Names. Unusual proper nouns get mangled, and the model has no way to learn. A novel set somewhere with unfamiliar place names will mispronounce them consistently for the whole book.
Emotional register. It reads a death and a shopping list identically. For literary fiction this is the thing that eventually wears you down.
Poetry. Line breaks carry meaning that the model cannot see.
If you are choosing between synthetic narration and a well-made commercial audiobook of the same title, the commercial one is better and it is not close. This is for the enormous number of books that have no audio edition at all.
Which voice
There is one engine and eleven voices, so the only choice is who reads to you.
Six are American, four English and one Scottish, in both male and female. There is no Australian voice — the model’s publisher has not released one — and the Scottish voice is female only. Saying so is more useful than letting you hunt for something that is not there.
The differences between them are larger than you would expect from a list of names. Some are warmer, some crisper, and a voice you find pleasant for twenty minutes is not always one you can live with for eleven hours. Sample two or three.
We measure your machine on a few paragraphs before you commit and show you the real estimate, so the decision to start is made against a number rather than a guess. It is usually about an hour for a ten-hour book, on the processor you already have.
Pacing
A separate setting from speed, and more useful than it sounds. It controls the silence between sentences, paragraphs and chapters rather than how fast the words come out.
Natural suits most books. Tight removes some of the air, which suits dense non-fiction you are working through rather than enjoying. Relaxed lengthens the pauses, which helps with anything you want to think about between paragraphs.
If you want the words themselves faster, do it in your player. Every audiobook app has a speed control, and adjusting it there means you can change your mind without converting again.
Practical advice
Convert a short book first. Twenty minutes of something you know well will tell you more about whether you can live with the voice than any description.
And listen to the sample before starting an eleven-hour job. The preview button exists precisely because an hour is a long time to discover you dislike a narrator, and eleven hours of listening is a great deal longer.