Text to Speech for Audiobooks

Novels and essays are written to be read silently. Narration needs them rewritten, and this is the single largest quality difference between a good audiobook and a robotic one.

Parenthetical asides are the main offender. An aside is invisible on a page and a stumble in audio, because the narrator has to hold the main clause, deliver a subordinate thought, and return. Move it into its own sentence.

Dialogue tags and em-dash interruptions have the same problem in fiction. So do footnotes, symbols, and numbers written in digits. Resolving all of these in the source text is faster than trying to fix them by regenerating, because the model reads what is written rather than what is meant.

A single audiobook chapter is usually 30,000 to 60,000 characters, so it is fifteen to thirty jobs. That is not a workaround, it is the shape of the work, and treating it as one is where projects stall.

Split at scene boundaries. A scene break gives you sections of three to eight thousand characters, each generating as two or four jobs, and it means the splits are somewhere that already made structural sense. A character-count split at 2,000 puts a seam mid-sentence.

Name the files as you go. Thirty sequential files joined in the wrong order is a genuinely painful mistake to discover at the end, and the numbering costs nothing at the start.

Pacing is the difference between a listener finishing a chapter and abandoning it at fifteen minutes, and it is mostly the paragraph pause setting rather than the speed.

A longer paragraph pause at the end of a section reads as a beat between scenes, which is exactly what a human narrator does. At the default of 500 ms the joins between generated jobs feel abrupt, and over an hour that accumulates into a sense of breathless delivery.

Keep speed constant across the whole chapter. A listener who has adjusted to one pace finds a change jarring in a way that a single odd voice never does, because the speed is what they calibrated to without noticing.

Fiction and non-fiction want different voices more than most people expect. Fiction wants a voice that disappears, so the reader stops hearing narration and starts hearing the story. Non-fiction wants a voice that carries the argument, because the words are doing the work rather than the delivery.

Multi-character fiction is the honest limitation. There is no speaker labelling, so distinct voices have to be produced by generating each character separately, which is workable for a two-hander and impractical for an ensemble cast.

For first-person narration, match the voice to the character you have written. A younger, higher voice for a younger narrator and a lower one for an older narrator are available in the catalogue, and the fit is more noticeable than the choice of accent.