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. One architectural question is where that computation runs; model design and the surrounding application matter too.

• 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. Once the required model is installed, this speech step does not need to upload the text or receive audio from a hosted speech service.

This describes the speech step, not every feature of an app. Downloading a model, fetching a book, syncing files, or sending a message can still use a network. Check which steps process your content locally and which services receive it. We cover that distinction in private text to speech with no cloud.

Earlier systems often used recorded speech fragments or hand-tuned acoustic rules. Their joins and intonation could sound mechanical, although quality varied between systems. Local speech is not one fixed sound: it describes where generation happens, not a particular synthesis technique or a ceiling on quality.

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 was the computational cost of generating samples one at a time. Later models and faster hardware made neural speech easier to deploy on everyday devices. WaveNet is part of that history, not a benchmark for the voices in today's apps.

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 processors can accelerate suitable models. On Apple platforms, Core ML offers CPU, GPU and Neural Engine options. The runtime and supported operations determine where the work runs; having an accelerator does not remove every bottleneck.

A suitable model and runtime can generate speech without a remote server. LoudReader uses local speech for book narration. Other local tools, including Loudkit, download models during setup and then generate speech locally. Packaging varies by product and release; check resource availability before relying on disconnected use. Apple documents the runtime choices in its Core ML compute-unit guide.

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. They may come with the app or arrive in a separate download. Allow room for the model, generated audio, and later updates.
• Voice variety. A cloud service may offer a different voice and language selection from a local tool. LoudReader offers 23 studio narrators across 10 languages, subject to device capability; Loudkit has its own 28-voice collection. Check the actual roster and samples rather than assuming the two products expose identical choices.
• Hardware floor. Real-time speech performance depends on the model and runtime. LoudReader requires iOS/iPadOS 18+ and can run its iPad build on compatible Apple Silicon Macs with macOS 15+. It has no separate native Mac build. Its requirements are not a rule for every local speech tool.

Local synthesis avoids sending each passage to a hosted speech provider and keeps working without that provider's connection. Pricing, quotas, accounts, file sync and diagnostics are separate product decisions. LoudReader 1.12 also includes Sentry diagnostics and TelemetryDeck usage analytics, enabled by default; the privacy toggle is not exposed in that release. An on-device engine alone does not make an entire application network-free. See the privacy policy.

Where a model runs does not give it a quality score. Compare the voices on your own material: dialogue, unfamiliar names, numbers, and a longer passage. For reading, the useful question is whether you can follow a chapter comfortably. You can install LoudReader, prepare your chosen voice and book, disconnect and try a new passage. 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.

LoudReader is the app for listening to your books and documents. Loudkit is the Apache-2.0 open-source speech framework for developers who want to add a voice to their own software. It provides SDKs and a command-line tool; you supply the application around it.

Loudkit for agents is a separate developer preview for connecting local speech to an existing agent. Its current voice service needs an Apple Silicon Mac. Speech synthesis and transcription run locally after setup, while the agent's model and messenger can still use online services. The setup guide distinguishes tested speech providers from messenger delivery that still needs validation.