Text-to-speech has been around for decades; narrating an entire novel so that it is pleasant for ten hours is a newer problem. This is how it is done and what to expect.
From file to sentences
The book is parsed into chapters and paragraphs (see the format guide), then split into chunks that a speech model can voice in one go: a few sentences to a paragraph. Chunk boundaries follow sentence and paragraph breaks so a breath never lands mid-clause. Chapter titles are voiced as their own short cue.
From sentences to speech
Modern neural speech models generate audio directly from text and a voice identity, predicting pitch, pace and emphasis from context rather than stitching recorded syllables. That is why current voices can carry a question, a sarcastic line or a list without sounding flat. Book2Speech narrates with licensed voices from a commercial provider; the catalogue spans 260+ voices in 40+ languages with gender, age and accent variety.
Why long-form is different
•
Continuity: consecutive chunks must sound like one performance, so voice settings stay fixed across the book and chunk edges are joined without gaps or overlaps.
•
Latency: nobody wants to wait for a ten-hour render. Narration is generated just ahead of the listener and streamed, with a buffer that refills as you go.
•
Cost: generating audio is far more expensive than serving it, so generated chapters are cached per voice and replayed for free.
•
Structure: front matter, part dividers and back matter need to be recognised so “start listening” lands on chapter one.
Where it still falls short
•
Unusual names and invented words are pronounced by best guess; there is no per-book pronunciation editor in Book2Speech yet.
•
Heavy dialogue between many characters is read in one voice; there is no automatic multi-voice casting.
•
Poetry, tables and equations are read as text, which is often not what you want.
•
OCR errors in scanned books are read faithfully.
Human narrators versus AI
A good human narrator is still the gold standard for a performance. AI narration wins on availability: the book you have, in the voice you choose, today, in a language where no recorded audiobook exists. Most people use it for exactly the books that will never get a studio recording: drafts, papers, textbooks, small-press and out-of-print titles, and classics in translation.