Why offline AI voices matter

A novel is eight to twelve hours of speech. Roughly half a million characters. That number is the whole argument, so it’s worth sitting with before anything else.

1. The cost stops being invisible

Cloud text-to-speech is billed per character. At book scale, that isn’t a rounding error: it’s why every service built on it sells a subscription rather than a purchase. You can pay monthly for as long as you keep reading, or you can run the model on hardware you already own and pay nothing per hour.

This isn’t an argument that subscriptions are wrong. It’s an argument that reading, of all activities, shouldn’t have a meter attached to it.

2. Your books are not anonymous data

To speak your text, a cloud service must receive your text. For an article that’s trivial. For a library it’s a record of everything you read, in order, at what time of day, and how far you got. Reading history is unusually revealing: health, faith, politics, grief, whatever you were quietly researching last month.

And it isn’t only books. People run contracts, medical letters and unpublished manuscripts through read-aloud tools. On-device narration means that content never leaves the phone, which is a guarantee no privacy policy can match.

3. It works where you actually read

Underground trains, aircraft, rural roads, hotel Wi-Fi that requires a login every ninety minutes. A voice that lives on your device starts in the same second whether you have five bars or none. A cloud voice buffers, stutters at the edge of coverage, and stops entirely at 30,000 feet.

There’s a subtler version of this: an on-device engine can generate the next few paragraphs ahead of time, so a chapter break costs nothing. Nobody notices when this works. Everybody notices when it doesn’t.

4. Battery and data, with the nuance included

The honest version: generating speech locally uses the processor, and streaming it uses the radio. Neither is free. In practice, sustained network use tends to cost more battery than a small, efficient model does, and on-device narration uses no mobile data at all, which for long listening is the bigger number.

Model weight matters here. A light engine like Piper barely registers; a heavier one like Kokoro asks more of the chip and, on an older phone, will show up in battery life. Choosing between them is a real decision, which we cover in Piper vs Kokoro.

5. It keeps working

Cloud services change. Prices rise, free tiers close, APIs get deprecated, companies are acquired. A voice model sitting in your app’s storage has none of those failure modes: it will read to you in five years exactly as it does today, whether or not anyone is still maintaining the service that inspired it.

For a reading app, which people keep for years, that stability is worth as much as the sound quality.

What changed, technically

None of this was possible until recently, and it’s worth being precise about why. Neural speech models used to need a data centre. Two things fixed that: architectures that produce good speech with far fewer parameters, and phone chips that got genuinely fast at this kind of work.

Piper came out of the open-source voice-assistant world, with small per-voice models and an emphasis on speed and language coverage. Kokoro went the other way, spending more capacity on how human the result sounds while staying small enough to run on a phone. Between them they cover most of what a reader needs, and both are open, which is why an independent app can ship them at all.

The honest counter-arguments

• The best cloud voices are still better. For expressiveness, character voices and emotional range, a large hosted model wins. On-device has closed most of the gap, not all of it.
• Language coverage is uneven. Some languages have excellent on-device voices and some have none. Cloud services, with more resources, cover more ground.
• Models take space. Each downloaded voice occupies storage, which is a real cost on a 64 GB phone.

Where that leaves us: for a one-off article, use whatever sounds best. For your library, the voice should live where the books do.

What this looks like in practice

Lectern ships both engines, detects the language of the book you open, proposes a matching voice, downloads it once, and then never needs the network again. The sentence being spoken is highlighted so you can read along, the screen stays on while it narrates, and there is no account anywhere in the chain.

The point isn’t that offline is a feature. It’s that a book you own, read by a voice that runs on your own device, with nobody keeping a record, is what reading was before it became a service.

Frequently asked