Live Transcription

live-transcription

Vercel's Next.js Live Transcription (template: https://vercel.com/templates/next.js/nextjslive-transcription) — Deepgram's live-transcription starter: speak, and the page captions you live, each word as it is heard, over a visualiser of your microphone — running on Cloudflare Workers + Static Assets through @opennextjs/cloudflare, and packaged as an Orun baseline: one command takes a new transcription site from nothing to live on stage and prod, in its own repository, under its own name — in ONE landing. It transcribes from its first deploy, on Workers AI, with no key.

The template's page, contexts and components are the code. What the port adds is the Workers build, the Worker's own entry, the monorepo wrapper, a test package and the deploy contract; what it changes is what Workers, and time, force:

• The service: the template needs a Deepgram API key: its /api/authenticate route mints a short-lived token with it, and the browser streams the microphone to Deepgram's own WebSocket. Here, with no key, the page streams to the Worker instead: GET /api/listen (apps/web/worker.ts → lib/listen.ts) opens Deepgram's Nova-3 on Workers AI over a WebSocket through the Worker's own AI binding (env.AI.run("@cf/deepgram/nova-3", …, { websocket: true })) and hands that socket to the browser. Nova-3 sends Deepgram's own Results messages, so the template's caption code reads them unchanged. With the owner's optional Deepgram key, the route mints a token and the browser streams to Deepgram directly, as upstream did.
• The audio: Nova-3 on Workers AI takes linear16 PCM, so an AudioWorklet (public/pcm-capture.worklet.js) taps the microphone's stream at the AudioContext's own rate (lib/workers-ai-live.ts); Deepgram's own socket keeps the MediaRecorder's WebM chunks, as upstream sends them.
• Next.js 14 → 15.5: OpenNext 1.20.16 builds Next 15 from 15.5.29, so the template moves there, and to React 19 (two type-level changes in its own code).

What the baseline adds is everything needed to make a copy of it somebody else's — built the way cirrus-baseline is. See specs/epics/live-transcription for the design, the milestones and the as-built status.

Instantiating a site

git clone --depth 1 --branch <baseline-tag> https://github.com/sourceplane/live-transcription
cd live-transcription
orun new --blueprint repo-blueprint.yaml --out ~/apps/acme-notes --run-hooks --resume \
  --set reponame=acme-notes --set productname="Acme Notes" \
  --set githuborg=acme-inc --set subdomain=<workers-dev-subdomain>

Or, once the platform's registry lists it, from the Orun console (Baselines → live-transcription → Build). Two phases — 01-product (the repository, its CI and the whole app in one pull request, deployed to stage and prod by one convergence, smoke-tested) and 02-docs (the live deployment recorded) — each landing one pull request in the new repository. BOOTSTRAP.md is the operator guide; docs/phases/ has a page per phase.

The new repository receives the app only — the app, its tests, its CI, its tooling, a README of its own and its operating contract — renamed throughout to the app's name, repository and Cloudflare Worker. None of this repository's machinery ships, and nothing that ships names it. The upstream template's own names — its repository and Vercel page, which end in this baseline's slug — are left as they are, and so are Deepgram's SDK names.

The only provider the bootstrap asks for is Cloudflare: Workers AI is a binding on the same account, with nothing to provision. Deepgram is not a provider the platform brokers, so the app is built without a key, and its owner may store one afterwards — one optional Orun secret, DEEPGRAM_API_KEY, and a redeploy, which pushes it to the Worker as a secret. The app's README, "Use your own Deepgram key" (.baseline/product/README.md here), has the steps.

What is here (the site)

Getting started

pnpm install
pnpm typecheck
pnpm lint
pnpm test
pnpm build            # next build + opennextjs-cloudflare build → apps/web/.open-next

That much needs no Cloudflare account and no key: the baseline's own e2e tier (testing/e2e.test.sh) runs exactly this on each landing before a tag, and an app's developers run it before a push — an app's pull request builds and dry-deploys it, and does not lint or run the tests. Workers AI has no local mode, so to transcribe on your machine either log in to Cloudflare and preview the built Worker, or give next dev a Deepgram key:

cd apps/web
npx wrangler login                # Workers AI runs on your account, even in wrangler dev
pnpm build && pnpm preview        # wrangler dev over the built Worker (worker.ts and all)
# — or —
cp .env.example .env.local        # DEEPGRAM_API_KEY
pnpm dev                          # http://localhost:3000, transcribed by Deepgram directly

How it maps to Vercel

CI