SingSpree
Free online karaoke with friends. Pick a song, start a party, and sing together in real-time with pitch scoring.
SingSpree
Features
- 20,000+ songs with synced lyrics and pitch targets
- Real-time multiplayer -- host a party, share the code, sing together
- Pitch scoring -- client-side pitch detection (ONNX model via Web Audio API) with server-authoritative scoring
- Live note visualization -- see your pitch and other players' notes in real-time on a scrolling music bar
- Song queue -- drag-to-reorder queue so the party keeps going
- Share cards -- Spotify-Wrapped-style score cards you can share to social media
- 21 languages -- full i18n; the language is a client-side preference, not part of the URL (one canonical URL per page)
- Mobile-friendly -- responsive layout with portrait and landscape support
- Auto-skip -- optionally skip intros, outros, and non-music sections
- Gap correction -- drag-to-fix timing offset for any song
Tech Stack
- React 19 with React Router 7
- Tailwind CSS v4 -- dark neon theme, no component library
- Vite 8 -- dev server, build, and HMR
- Web Audio API -- real-time pitch detection via AudioWorklet + ONNX Runtime (WASM)
- WebSocket -- live multiplayer sync (notes, scores, queue, player state)
- i18next -- 21 locales with browser language detection
- Nginx -- production serving; bots get prerendered HTML from the backend (
/prerendered volume), legacy /{lang}/… and slug URLs get 301s
Getting Started
Requires Bun >= 1.2.
bun install
bun run dev
Dev server starts on localhost:3001, proxying API requests to the backend at localhost:3000.
Scripts
| Command | Description |
|---|
bun run dev | Start Vite dev server |
bun run build | Production build to build/ |
bun run preview | Preview production build locally |
bun run test | Run tests (Vitest) |
bun run test:e2e | Playwright end-to-end tests (needs the backend on :3000) |
bun run test:stress | Many-singer stress test, see below |
Stress test
bun run test:stress puts a dozen singers into one party: real Chrome contexts
with a fake microphone, each running the whole pipeline (AudioWorklet → ONNX
pitch worker → WebSocket), the first one hosting. Every browser carries an
in-page monitor, and the run ends with a per-client table (frame rate, long
tasks, JS heap, notes in/out, latency) plus assertions that nobody
disconnected, everyone is on the scoreboard, the singers' notes reached the
host and the host stayed above MIN_FPS. Past eight singers the server puts
only eight on screen (three top scores pinned, five lanes rotating one per
lyric line), so from then on the host hears the lanes rather than everyone,
and the run prints who they were.
| Env | Default | Meaning |
|---|
REAL_SINGERS | 12 | Chrome contexts (incl. the host) |
BOT_SINGERS | 0 | extra synthetic singers driven from Bun over WebSocket (e2e/wsSinger.js), cheap enough for dozens |
STRESS_SECONDS | 30 | how long everyone sings |
MIN_FPS | 20 | host frame-rate threshold |
CPU_THROTTLE | 1 | DevTools-style CPU slowdown for every real browser: 4 ≈ mid-range phone, 6 ≈ an old low-end one |
Serve the production build for realistic numbers (bun run build && bunx vite preview;
the test reuses whatever is on :3001), because React's development runtime
alone dominates a profile. Emulating an old phone with a dozen singers:
REAL_SINGERS=1 BOT_SINGERS=11 CPU_THROTTLE=6 bun run test:stress.
Mind what you are measuring: every real context runs its own ONNX pitch
worker and YouTube player, so a dozen of them on one laptop saturate the CPU
and every client (host included) stalls — the run then prints a warning that
the host video stopped advancing. That measures the machine, not the app. To
measure the frontend's cost of rendering many singers, keep the real
browsers few and add bots: REAL_SINGERS=2 BOT_SINGERS=40 bun run test:stress.
Docker
docker build -t singspree-fe .
docker run -p 80:80 singspree-fe
Multi-stage build: Bun 1 Alpine for bun install --frozen-lockfile && bun run build, then Nginx Alpine to serve the static files. The build stage is deliberately not cached in CI -- installing takes about eight seconds, and caching it meant pulling a 577 MB node_modules layer out of the GitHub Actions cache first.
Project Structure
src/
components/ UI components (MusicBars, Lyrics, ShareCard, ...)
logic/ Pitch detection, mic input, lyrics parsing, WebSocket handling
pages/ Route pages (Entry, Party, Join, compliance pages)
i18n/locales/ 21 locale JSON files
index.css Tailwind v4 theme (neon-cyan, neon-purple, neon-magenta, neon-green)
License
MIT
