Hosting and deployment
What your pages need to serve the SDKs, how to self-host models for strict CSP or offline use, and how bundlers pick up the workers.
Requirements
- HTTPS (or
localhost) for the microphone. - WebGPU for text-to-speech. Nothing to configure, but check with
detectCapabilities()and explain it to users whose browser lacks it. - Cross-origin isolation is optional. It makes the WASM fallback multi-threaded. If you enable it, use
Cross-Origin-Embedder-Policy: credentialless, notrequire-corp, which would block model downloads from Hugging Face:
Cross-Origin-Opener-Policy: same-origin
Cross-Origin-Embedder-Policy: credentialless
Static hosts that can’t set headers, like GitHub Pages (which serves this site), work fine without it.
Bundlers
Vite, esbuild and Bun each take one plugin; webpack 5 and Next.js need none. See Bundlers.
Self-hosting models
By default models come from Hugging Face and are cached in the browser’s Cache Storage. For a strict Content-Security-Policy, an intranet, or full offline control, mirror the repositories and point the engines at them:
import { createSTT } from "@kucukkanat/stt";
import { createTTS } from "@kucukkanat/tts";
const transformers = {
remoteHost: "https://models.example.com/", // serves <repo>/resolve/main/<file>, like huggingface.co
wasmPaths: "https://static.example.com/ort/", // ONNX Runtime's .wasm/.mjs files
};
const tts = createTTS({ transformers });
const stt = createSTT({ transformers });
The repositories are listed on each model (TTS_MODELS[key].repo, STT_MODELS[key].repo).
Content-Security-Policy
worker-src 'self' blob:: the engines’ workers, plus the microphone’s AudioWorklet (loaded from a Blob URL). Ifblob:isn’t allowed, servemicWorkletSource()yourself and passworkletUrltoopenMicrophone().connect-srcmust include where models come from (https://huggingface.coand its CDN, or yourremoteHost).script-src 'wasm-unsafe-eval'for ONNX Runtime’s WebAssembly.