Which AI voice models can your device run?
In-browser text to speech ranges from a 57 MB model any phone can handle to a 1.5 GB flagship that needs a real GPU. This page reads your browser's capabilities — locally, in about a second — and gives a verdict for every model in our lineup.
Browser support for WebGPU (mid-2026)
| Browser | WebGPU | Notes |
|---|---|---|
| Chrome / Edge | ✅ Since v113 (2023) | Windows, macOS, Android 121+; Linux still partial |
| Safari | ✅ Since v26 (Sept 2025) | macOS 26 / iOS 26 and later; older Apple OS versions have none |
| Firefox | ✅ Since v141 (July 2025) | Windows first, then macOS on Apple Silicon; Linux/Android pending |
| Android (Qualcomm GPUs) | ⚠️ Partial | WebGPU works but fp16 shaders are unavailable on Adreno — full-precision models only |
Every model also has a WebAssembly CPU path where feasible — see each model page for specifics.
Frequently asked questions
What is WebGPU and do I need it?
WebGPU is the browser API that lets web pages use your graphics card for computation. Small TTS models (KittenTTS, Piper, Kokoro) run fine on CPU without it; large models like Chatterbox effectively require it. Chrome and Edge have shipped it since 2023, Safari since version 26 (2025), and Firefox since 141 on Windows.
Why does my Android phone fail on some models?
Most Android GPUs (Qualcomm Adreno) don't expose fp16 shader support, so GPU-optimized half-precision models can't load. The checker detects this and points you to models with full-precision or CPU paths instead.
Does the check upload anything or install anything?
No. It reads capability flags your browser already exposes (GPU adapter, CPU core count, memory hints) locally, in about a second. Nothing is sent anywhere and no model is downloaded until you choose to run one.
My device says 'slow' — can I still use TTS?
Yes. 'Slow' means slower than real-time for that specific model. Pick a lighter tier: Piper synthesizes ~9× faster than real-time on almost any CPU, and KittenTTS runs on virtually anything with its 57 MB download.