README.md

metadata

title: Browser Voice Agent Demo
emoji: 🎙️
colorFrom: indigo
colorTo: purple
sdk: static
app_build_command: npm run build
app_file: dist/index.html
custom_headers:
  cross-origin-embedder-policy: require-corp
  cross-origin-opener-policy: same-origin
  cross-origin-resource-policy: cross-origin

Browser Voice Agent Demo

Browser-first, on-device voice agent demo with Gradio-style camouflage. The app is served locally with Vite during development and built as static assets for public deployment.

Stack

• Vite for local development, preview, and bundling
• Bun for package management and tests
• @mlc-ai/web-llm for local browser LLM inference
• @huggingface/transformers for local ASR
• Vendored Pocket TTS browser worker derived from KevinAHM/pocket-tts-web

Local Development

bun install
bun run dev

Then open http://localhost:3000.

The same dev server now carries the isolation headers, so Pocket can take the threaded path without a separate custom preview server.

For a production-like static serve, use:

bun run preview

That serves the built dist/ output with the same COOP/COEP/CORP headers.

Test

bun test

Build

bun run build

The production output is emitted to dist/.

Pocket TTS Runtime

The app now ships a built-in Pocket TTS browser runtime and same-origin model assets:

• src/vendor/pocket-tts/inference-worker.js
• src/vendor/pocket-tts/sentencepiece.js
• src/vendor/pocket-tts/CODE-LICENSE
• src/vendor/pocket-tts/onnx/*
• src/vendor/pocket-tts/tokenizer.model
• src/vendor/pocket-tts/voices.bin
• src/vendor/onnxruntime-web/*

Those files are copied into dist/ during bun run build and served locally from:

• /pocket-tts/inference-worker.js
• /pocket-tts/sentencepiece.js
• /pocket-tts/onnx/*
• /pocket-tts/tokenizer.model
• /pocket-tts/voices.bin
• /onnxruntime-web/*

The vendored worker code is Apache 2.0; its license file is preserved in src/vendor/pocket-tts/CODE-LICENSE.

For best Pocket performance, serve the app with:

• Cross-Origin-Opener-Policy: same-origin
• Cross-Origin-Embedder-Policy: require-corp

Without those headers, the worker still runs, but ONNX Runtime drops back to single-threaded WASM.