On-Device Text to Speech, Explained

Every text-to-speech system does the same job: it turns written words into an audio waveform that sounds like someone saying them. The only architectural question is where that computation runs.

• Cloud TTS: your text is uploaded to a server, a large model there generates audio, and the audio streams back. You need a connection, and your text, whatever it is, gets processed on hardware you don't control.
• On-device TTS: a compact model stored on your machine does the same generation locally. Nothing is uploaded, nothing streams, and switching off Wi-Fi changes nothing.

The distinction sounds technical but the consequences are practical: offline reliability, zero data cost, and privacy that is structural rather than promised. A cloud service protects your documents with a policy; an on-device app protects them with the absence of a server. That difference is the entire subject of private text to speech with no cloud.

Because the techniques that fit on a device were the ones that sounded mechanical. Classic offline synthesis either stitched together tiny fragments of recorded speech (concatenative synthesis) or generated sound from hand-tuned acoustic rules (formant synthesis). Both were small and fast; both produced the flat, seam-riddled voice everyone associates with “computer reading.”

The turning point was neural speech synthesis. DeepMind's WaveNet paper (2016) showed a neural network generating the raw audio waveform directly, and it substantially narrowed the gap to human-sounding speech. The catch: that first generation of neural TTS was computationally enormous. It lived in datacenters because nothing else could run it. So for years the market split, natural voices in the cloud and robotic voices on the device, and every reading app had to pick a side.

Two curves crossed, from opposite directions:

• Models got smaller. Researchers spent the years after WaveNet making neural synthesis efficient. New architectures generate speech in far fewer steps, and compact open models now produce natural prosody at a fraction of the original computational cost.
• Devices got ML hardware. Phone and laptop chips stopped being general-purpose only. Apple Silicon ships a dedicated Neural Engine, silicon whose whole job is running neural networks, alongside fast unified memory that lets a model stream data without bottlenecks.

When a small-enough model meets a fast-enough accelerator, the datacenter stops being necessary. LoudReader is a concrete example of the pattern: its voice models ship inside the app, execute on the Neural Engine via Core ML, and the engine is written so it cannot download anything at runtime. Offline isn't a mode here, it's the architecture. That's what “natural offline voices” means mechanically: a neural model that once needed a server rack, running quietly on the chip you already own.

Honesty matters more than cheerleading here, because the trade-offs are real. They're just different from what people expect:

• App size. The models have to live somewhere. An on-device app ships its voices inside the download, so it's heavier than a thin cloud client. You pay in storage once instead of in bandwidth forever.
• Voice variety. A cloud service can host thousands of voices in dozens of languages because they all live server-side. An on-device app ships a curated set. LoudReader offers 8 natural offline voices, and its voices are English-only today. If you need Mandarin or a celebrity soundalike, the cloud genuinely serves you better.
• Hardware floor. Real-time neural synthesis needs a modern chip. LoudReader requires macOS 15+ on Apple Silicon, or iOS 18+ on iPhone. An Intel Mac can't run it, and pretending otherwise would help nobody.

What you get in exchange: no word quotas metered by a server, no subscription tied to compute costs someone else is paying, no outage that takes your reading with it, and no copy of your documents anywhere but your own disk.

The gap has narrowed dramatically, and it keeps narrowing, but the honest answer is that the largest cloud models are still ahead on expressiveness and variety, because scale still buys something. The more useful question for a reader is whether an on-device voice is natural enough to disappear into a book for three hours. That bar has been cleared, and it's a claim you can test in minutes without trusting anyone: install LoudReader, switch off Wi-Fi, and play a chapter. The Mac-specific picture, including what the built-in system voices can and can't do, is covered in offline text to speech on Mac; the iPhone side is in text to speech without internet on iPhone.