Apppublic

Riksarkivet/multimodal-webgpu-demo

sourceHugging Faceapache-2.0updated 6mo agoView on Hugging Face

2likes

App README

Multimodal Lab

A local-first, multimodal playground for browser ML. Runs KB-Whisper (Swedish) and OpenAI Whisper (English) directly in the browser via Transformers.js v4 on WebGPU. Nothing leaves the tab — audio, model weights and transcripts all stay on the user's machine.

Routes

—/ — Realtime speech: live transcription from microphone or tab/system audio, with on-Stop finalisation for word-accurate timestamps.
—/batch — Batch media: drop in a stack of audio/video files, chunked transcription of any length, downloads per file.
—/search — Multimodal search (WIP).
—/about — Privacy/tech overview.

Stack

—SvelteKit 2 + Svelte 5 (runes) · Tailwind v4 · shadcn-svelte · bits-ui
—@huggingface/transformers v4 · ONNX Runtime Web
—@sveltejs/adapter-static — pure SPA, no server

Develop

bash

bun install
bun --bun run dev

Build

bash

bun run build

Produces a self-contained static site in build/.

Deploy to Hugging Face Spaces

This repo is configured as a static Space. First-time setup:

bash

# 1) Build the site
bun run build

# 2) Create a new Space (one time), then clone its git remote
git clone https://huggingface.co/spaces/<your-org>/<space-name>
cd <space-name>

# 3) Copy the contents of build/ (not the folder itself) into the Space repo
cp -r ../audio-demo/build/* .
cp ../audio-demo/README.md .   # keep the YAML frontmatter above at the top

# 4) Push
git add -A
git commit -m "Initial deploy"
git push

Subsequent deploys: re-run the build, copy build/* into the Space repo, commit, push.

Model weights are not included in the build — Transformers.js fetches them from the HF Hub on first run and the browser caches them for later visits.