TTS/STT Guide (step by step)

This page tests the offline, client-side voice engine — nothing here talks to a cloud API or this site's server once the model is loaded. Everything below runs in your browser.

Step 0 — What your browser supports

This checks WebAssembly/SIMD, storage, and file/microphone APIs before asking you to download anything. If a row above shows ❌, the step that depends on it will say so.

Step 1 — Get the voice model

Download English voice — en_US-amy-low (~18MB)

This is a real piper voice (bundled via sherpa-onnx's model zoo), fetched by scripts/fetch-tts-assets.sh. Save it somewhere you'll find again — Step 2 loads it back from disk, not from this site.

(The inference engine itself isn't wired up yet — see the note in the page source. Steps 2-4 below are placeholders until that's built.)

Step 2 — Load the model

Two ways, depending on Step 0's results:

• If your browser supports the File System Access API: click "Load model," pick the file once, and it can be reloaded automatically on later visits.
• Otherwise: click "Load model," pick the file each time you visit — still fully offline, just one extra click.

(Loader control placeholder — added alongside Step 1.)

Step 3 — Try it: Listen (TTS)

(Same "Listen" control as the earlier Web Speech demo, once wired to the loaded model instead of the browser's built-in — see the English sample for the older browser-native version for comparison.)

Step 4 — Try it: Speak (STT)

(Requires microphone access — Step 0 will show whether your browser supports this at all before you try.)

---

See design notes for how this approach was chosen over the browser-native Web Speech API.