Cloud TTS Apps vs On-Device: Why Credits and Caps Exist
Ever wondered why text-to-speech apps meter your listening with credits and stop working offline? It comes down to where the voice runs. Here’s the economics — and the alternative.
The question nobody answers
Install a popular text-to-speech reading app and you will quickly hit the same walls: a monthly cap on listening hours, "premium voice" credits that run out mid-chapter, and an app that goes silent the moment your phone loses signal.
None of that is malice. It is architecture.
How cloud TTS works
Most reading apps do not synthesize speech on your phone. When you press play, your text is sent to a server, a large voice model renders it to audio, and the audio streams back. That design has real consequences:
- Every minute you listen costs the company server money. -- GPU time is expensive, so heavy listeners are unprofitable — unless usage is capped or metered with credits.
- No connection, no voice. -- The model lives in a data center. On a plane, in a subway, in a rural area, the app simply cannot speak.
- Your reading goes through their servers. -- Whatever you listen to is, at minimum, processed by someone else's infrastructure — and subject to their content rules.
If you have ever felt that billing in this category is confusing or distrustful, this is the root cause: the pricing has to recover a per-minute cost you never see.
How on-device TTS works
The alternative is to run the voice model on the phone itself. Modern compact neural models like Kokoro make this possible: they are small enough to run on everyday Android hardware, yet natural enough for hours of listening.
On-device flips every one of the economics above:
- Listening costs the developer nothing extra -- so there is no reason to meter you. Unlimited actually means unlimited.
- Offline is the default, -- not a missing feature. The voice is already on your phone.
- Your books never leave your device. -- Nothing is uploaded, logged, or filtered.
The honest trade-offs
On-device is not magic. The very largest cloud voices, running on data-center GPUs, can still sound marginally more expressive than a compact on-device model. And the first voice download takes a little storage on your phone.
But for long-form listening — actual books, not ten-second demos — a well-tuned on-device engine is more than good enough, and the freedom from meters, caps, and connectivity is transformative.
Where Echo fits
Echo was built on exactly this bet: an on-device Kokoro engine, extensively customized for multi-hour listening, in the free tier. No credits, no hourly caps, no upload. The optional Premium tier sells signature narrator voices — not the basics.