Technical Guide
How On-Device Text-to-Speech
Works on iPhone
Updated October 2026 · Neural Engine, AVSpeechSynthesizer, and voice models
When you hear your iPhone read text aloud, one of two things is happening: either the text is being sent to a cloud server for processing and audio is streamed back, or the entire speech synthesis pipeline is running on the device's hardware. This guide explains how on-device text-to-speech works — the hardware, the software, and why it matters.
The hardware: Apple's Neural Engine
Every iPhone since the XS (2018) includes a Neural Engine — dedicated silicon designed for machine learning inference. The A12 Bionic started with 8 neural cores; current chips have 16+. This hardware runs neural network models at speeds that would be impractical on the CPU alone.
Text-to-speech is one of the workloads Apple has optimized for the Neural Engine. The TTS model runs on this dedicated hardware, converting text to waveforms locally at real-time speed. This is why on-device TTS can start instantly — no network round-trip, no server queueing.
The software: AVSpeechSynthesizer
Apple provides the AVSpeechSynthesizer framework for on-device text-to-speech. Developers create an AVSpeechUtterance with the text to speak, assign a voice, set the rate and pitch, and the framework handles everything else — text analysis, prosody generation, waveform synthesis, and audio output.
Under the hood, AVSpeechSynthesizer uses a neural network model for each voice. Apple offers three voice quality tiers:
- Default — Small model, always available, sounds robotic
- Enhanced — Larger neural model, must be downloaded (~150MB per voice). Natural-sounding with good intonation
- Premium — Highest quality neural model (~300–500MB per voice). Available on iOS 17+ for select languages. The most natural prosody and expression
LocalReader uses whatever voices you have installed on your device. The app recommends downloading Enhanced or Premium quality voices for the best experience.
The pipeline: from text to speech
1. Text analysis
The input text is analyzed for sentence boundaries, abbreviations, numbers, URLs, and special characters. "Dr. Smith arrived at 3pm" becomes "Doctor Smith arrived at three P M."
2. Prosody generation
The model determines pacing, emphasis, pitch contours, and pauses. Questions get rising intonation. Commas create brief pauses. The neural model has learned these patterns from human speech data.
3. Waveform synthesis
The neural voice model generates audio waveforms on the Neural Engine. This is the computationally expensive step — and the reason dedicated hardware matters. The result is a continuous audio stream.
4. Audio output + word callbacks
Audio plays through the device speakers or connected headphones. The framework sends real-time callbacks as each word is spoken — this is what LocalReader uses for word highlighting.
On-device vs. cloud: the trade-offs
| Dimension | Cloud TTS (ElevenLabs, Google, AWS) | On-Device TTS (Apple AVSpeechSynthesizer) |
|---|---|---|
| Voice expressiveness | Very high — large models, GPU clusters | Good — constrained by device hardware |
| Latency | 200–1000ms (network + processing) | Near-zero (local hardware) |
| Privacy | Text sent to external servers | Text stays on-device |
| Offline capability | None | Full |
| Cost model | Per-character or per-request | Free (built into the OS) |
| Voice variety | Hundreds to thousands | 30+ (depends on installed voice packs) |
| Languages | 50+ | 8–20 (depends on voices installed) |
Why on-device matters for reading apps
For a reading app like LocalReader, on-device TTS isn't just a privacy feature — it changes the fundamental user experience:
- Zero startup delay — Press play and speech starts immediately. No "loading" spinner.
- Consistent quality — Same voice, same speed, same behavior whether you're on Wi-Fi, cellular, or Airplane Mode.
- No usage metering — Cloud TTS charges by character. A 300-page PDF could cost dollars. On-device is free.
- Document confidentiality — Legal documents, medical records, personal journals — nothing leaves the device sandbox.
- Battery efficiency — No radio usage for network requests means less battery drain during long listening sessions.
FAQ
What is on-device text-to-speech on iPhone?
On-device TTS uses neural speech models that run directly on the iPhone's Neural Engine — a dedicated hardware chip for machine learning. Apple's AVSpeechSynthesizer framework provides this capability. Text is processed on the device, speech is generated locally, and nothing is sent to any server.
How many on-device voices does iPhone support?
iPhone supports 30+ on-device voices across multiple languages. The number depends on which voice packs you've downloaded in Settings → Accessibility → Spoken Content → Voices. Apple's Enhanced and Premium quality voices use neural models for more natural speech.
Is on-device TTS as good as cloud-based TTS?
Apple's on-device neural voices are high quality with natural intonation and prosody. Cloud services (ElevenLabs, Google Cloud TTS) can offer more expressiveness and voice variety, but on-device voices have zero latency, work without internet, and keep your text private. For everyday reading, on-device quality is excellent.
Which iPhones support on-device neural TTS?
Any iPhone with an A12 Bionic chip or newer (iPhone XS, 2018 and later) supports on-device neural text-to-speech. Older devices fall back to basic system voices. LocalReader requires iOS 17+ which supports the latest Enhanced quality voices.
No account required · 100% offline · Free tier available