EarChords

Get the chords from any song — right in your browser. Upload a track (or play one into your mic), and EarChords runs a neural pitch‑detection model locally, recognises the chords, and builds a playable chord sheet with transpose, capo, tempo‑without‑pitch‑change, and guitar / piano / ukulele diagrams.

Privacy‑first: audio is decoded and analysed entirely on your device. Nothing is uploaded — there is no backend and no database. The build is a fully static site.

🔗 Live: earchords.com

How it works

1. Decode any audio to 22050 Hz mono in an OfflineAudioContext.
2. Detect notes with @spotify/basic-pitch (TensorFlow.js) — the model (~0.9 MB) lives in public/model/ and is lazy‑loaded on first analysis.
3. Recognise chords with a custom pipeline: chroma vectors → cosine‑similarity template matching → key‑aware priors + bass‑root and extension evidence → HMM/Viterbi global decoding for the most likely chord sequence.
4. Play along: <audio preservesPitch> gives tempo change without pitch shift; requestAnimationFrame keeps the chord sheet, waveform, and spectrum in sync.

Features

• Audio → chord sheet, 100% in the browser (audio never leaves the device)
• Key detection (Krumhansl‑Schmuckler) with sharp/flat spelling per key
• Tempo estimation (autocorrelation) and a numbered, bar‑based chord sheet
• Transpose · capo · slow down (50–100%) · follow‑scroll
• Guitar / piano / ukulele fingering diagrams, with click‑to‑cycle alternate voicings (barre chords included)
• A–B section loop (shift‑click two bars)
• Keyboard: space to play, ←/→ to step chords
• Export: copy as text, print‑friendly layout
• Light/dark themes, responsive

Develop

npm install
npm run dev        # http://localhost:5173
npm run build      # static output in dist/
npm run preview    # preview the production build
npm test           # music-theory / recognition regression tests

Deploy

npm run build produces a static dist/ you can host anywhere (Cloudflare Pages / Vercel / Netlify / your own Nginx). No backend, no database. Make sure dist/model/ (~0.9 MB) and dist/demo.wav ship with it, or first analysis and the demo button will fail.

The build also generates static, SEO‑friendly chord pages (songgen/) into clean URLs (e.g. /wonderwall-chords) plus sitemap.xml and robots.txt.

Project layout

• src/chords.js — chord recognition core (chroma cosine matching + key priors + segmentation + Viterbi + transpose)
• src/music.js — key detection (K‑S), tempo estimation, sharp/flat spelling
• src/audio.js — decode to 22050 Hz mono + Basic Pitch inference (lazy‑loaded)
• src/diagrams.js — guitar/piano/ukulele diagrams (SVG; hand‑calibrated open chords + movable/barre fallback + alternate positions)
• src/main.js — UI orchestration: bar sheet, playback, transpose/capo/tempo/loop/export/theme
• songgen/ — build‑time static SEO chord pages + sitemap assembly
• validation/ — accuracy validation (synthetic fixtures, regression tests, MIREX‑style scoring)

Accuracy & limitations

Accuracy is highest on synthetic audio and clean single‑instrument or singer‑songwriter recordings. Dense full‑mix pop is harder (a category‑wide problem) and may surface as N.C. or approximate chords. Automatic segmentation uses fine‑window recognition + HMM/Viterbi decoding + merging — it is not strict beat alignment. See validation/BENCHMARKS.md for benchmark results and the accuracy roadmap.