Three weeks of listening to an epub nobody will ever narrate

The book I tested with will never get an audiobook. It is a 900-chapter webnovel, translated by fans, abandoned by its original site, beloved by maybe four thousand people. Studio narration costs thousands of dollars per title; no publisher is coming. That is precisely the case where neural TTS matters, so I listened to it on my commute for three weeks, 40 minutes each way, and kept notes on everything that went wrong.

Week one: the voice was never the problem

Day one surprised me in the right direction. Modern neural voices breathe, pause, lift at questions. We later ran a small blind test in the office: five colleagues, one 2,000-word passage, human narrator versus TTS. Three of five guessed wrong. One of the two who guessed right said the giveaway was that the TTS never stumbled, which is a strange kind of compliment.

Day three, the actual problem introduced itself. Mid-scene, in a tense confrontation, a warm voice solemnly read out three asterisks. Then, two chapters later, “please vote for this chapter.” Wild epubs and TXT files are full of things that were never meant to be spoken: scene-break decorations, translator notes, promo footers, the occasional emoji. The engine reads whatever it receives. Text cleanup before synthesis is the single least glamorous feature in this whole pipeline, invisible in every screenshot, and it decides whether listening is pleasant or maddening.

Week two: chapter boundaries and the pocket test

Listening happens with the phone in a pocket and both hands on a subway rail. Every chapter boundary is a small trial: does playback roll straight into the next chapter, or stall while it synthesizes? Early in week two I got the stall version, a four-second gap that feels like a minute when you cannot reach your phone. The fix is prefetch: synthesize the next chapter while the current one plays. After that build, boundaries disappeared entirely, and honestly so did my awareness of chapters as a unit.

Caching matters for the wallet too. A chapter synthesized once should never be paid for twice, so the cache key wants to be chapter + voice + model. Re-listening is free; only an actual voice change re-bills. My three-week total for roughly 180 chapters came to a couple of dollars on a mid-tier voice, billed per character on my own key. The math: a 3,000-word chapter is 16,000-ish characters, a fraction of a cent to a few cents each depending on tier (rates are public).

Week three: one progress, two inputs

The habit that actually stuck was not “listening instead of reading.” It was switching. Ears for the steady middle chapters on the train, eyes for the lore-dense arcs at home in the reader. This only works because listening position and reading position are the same number; pause the narration at the station, and the page at home opens on that paragraph. In week one I had briefly used a setup with two separate positions. Every evening began with five minutes of scrolling, trying to remember whether my ears or my eyes had gotten further. One progress or no habit, that simple.

Voice choice turned out to be genre-dependent in a way I did not expect. The fast progression arcs wanted a young neutral voice at 1.2x. A slower historical interlude sounded wrong until I dropped to 1.0x with a deeper voice. There is no best voice, only a best voice for the next hundred chapters, which is an argument for apps that let you switch mid-book without re-billing the back catalog.

The tally

Three weeks, about 180 chapters heard, two real bugs (asterisks, boundary stall), one workflow lesson (shared progress), a couple of dollars spent. Fifteen years of commutes went to phone games before this. The book still is not finished, but for the first time a 900-chapter monster feels finishable, and not because I found more time. The time was always there. It just could not hold a book before.