On-device vs cloud narration: which to use, and when
There are two ways to turn a book’s text into speech, and they are good at almost opposite things. One runs on the chip already in your phone. The other runs on server hardware many times more powerful. Most apps pick one, hide the choice, and you live with the trade-off they made for you. It helps to understand both, because the right answer changes from book to book.
Why they sound different
The gap comes down to how big a model you can run and how much time you can spend.
On-device voices have to fit in a phone and speak in real time as you listen. That budget has grown a lot, and modern system voices are good, but it is still a budget. The model is small enough to run on battery, and it makes the audio on the fly, so it cannot plan far ahead for rhythm and emphasis. The result is clear and easy to follow, with faint seams between sentences and the odd flat line.
Cloud narration has neither limit. It runs a much larger speech model on GPUs, and it renders the book ahead of time, not in the moment, so it can shape the delivery across a whole sentence and paragraph. That is why produced cloud audio sounds smooth and deliberate where a device voice sounds merely competent. The cost: it takes compute, which is not free, and it has to be prepared before you listen rather than starting instantly.
What each one is actually best at
On-device narration is the quiet workhorse:
- Instant. No rendering, no waiting. Open a book and it plays.
- Free. No per-hour charge, no credits, no cap.
- Offline. Once a book is on your device it plays with no signal, on a flight or underground.
- Private. The text never has to leave your phone to be read.
Its one weakness is polish. On iOS you can close much of that gap by downloading the premium system voices, which sound noticeably better than the defaults. There is a short guide to that.
Cloud narration is the opposite balance:
- Richer. Produced, smooth, closer to a real recording.
- Multilingual at quality. Studio audio in BookNPC covers ten languages, rendered properly rather than approximated.
- Render once, keep forever. After it is made, it replays free and offline like any other file.
It costs two things: a moment to prepare, and a small amount per book. That is 25 cents per finished audio-hour in BookNPC, metered to the second, so a forty-minute book is around seventeen cents.
How to choose, per book
You do not pick one side for life. In BookNPC you can move a single book between on-device and cloud any time, even mid-chapter. A simple rule of thumb gets you the best sound where it counts without overspending:
- Start on-device for everything. It is free and instant, so there is no reason not to. A lot of listening, especially non-fiction you are in for the content, is perfectly good this way.
- Render studio for the books you love. When a book deserves to sound its best, a novel you are savoring or one you will relisten to, spend the few cents. You pay once and keep it.
- Use cloud when the language needs it. For produced audio in one of the ten supported languages, the cloud is the tool. On-device falls back to whatever voices your phone happens to have installed.
The mistake to avoid is rendering every book in the cloud by reflex. Most of your listening can be free on-device. Save the paid render for the titles where the extra polish earns its keep.
Why having both under one roof matters
A single-mode app forces a permanent compromise. An offline-only reader never sounds produced. A cloud-only one is useless on a plane, and it bills you for books you would happily have heard for free. Keep both, and the flight is covered by the free on-device voice while the favourite novel still sounds produced. You decide, book by book, which one is worth paying for.
The exact rates are on the pricing page. The short version is simple: on-device is the free default, cloud is the paid upgrade, and the switch is always yours.