Claude Rameau

"Yes — but why does it sound that way?"

Record the same riff on two guitars, drop the files in, and see why they sound different. Spectra, sustain, pickup voice, dead spots, overtone ring — measured, explained in a player's words, and never uploaded: the whole app is one HTML file that runs in your browser.

Open Claude Rameau → · press Load demo pair to see it work before you record anything.

What you get

One player, two guitars, the same phrase. You record the same riff on two of your own guitars — same picking, same amp, same cable, same room, same tuning — and drop one recording on each card. Because you and the phrase stay the same, what the plots show is the guitar.

• Before you compare ticks off what the two recordings share — register, note count, level, noise floor, recording path — whether the strings and scale differ, and whether the reliability bands are measured yet, so a difference below can be read as the guitar and not as the take.
• At a glance answers in sections, one line each: Sound (Single-coil — audibly brighter (1.00 kHz vs 474 Hz) · pickup voice peaks higher (3.70 vs 2.40 kHz)), Ring (Humbucker — overtones ring 1.4× longer · sustain 1.5× longer), the loudest gap, and what it adds up to for the player. Every item stands on a measured difference that cleared its own reliability band; the details sit one click away.
• Spectrum and difference — long-term average spectra with named regions, band energies, and a matched-EQ suggestion drawn on the face of a real device.
• Spectrograms and envelopes — with a harmonic overlay that predicts where each string's partials should land, and marks where two strings' harmonics collide.
• Tone character — pickup voice, body voice, overtone ring, bloom, sustain and dead spots, brightness, even/odd balance, pick attack — each with what it sounds like, what is measured, and how much the playing moves it. Click any value for the derivation with your numbers in it.
• Takes — record straight into a card or drop several files; every take counts, and the reliability bands narrow as takes accumulate.
• Share — a 1200×630 image of the comparison, or a PNG/CSV/JSON of any card.

Private by design — nothing you drop in ever leaves your browser

Your recordings are analysed right where they sit: in your own browser, on your own machine. Nothing is uploaded, and there is nothing to upload it to — no backend, no API, no CDN, no fonts or scripts fetched from anywhere. The one exception is a visitor counter (a Google tag) that loads only when you are online with Do Not Track off; it never sees your audio, your files, or anything you type.

That's not a policy you have to take on trust. It's the architecture:

• One file, no dependencies. Read index.html and you have read the whole program.
• Works with the network off. Pull the plug, open it from file://, and everything — including the demo pair, which is synthesised in the page — still works.
• Nothing is stored anywhere but your own browser. Your tuning, theme and colour choices live in localStorage; exports (PNG, CSV, JSON) are written straight to your downloads folder.
• Feedback leaves through your hands, not the page's. The Feedback button opens a GitHub issue form in a new tab with your setup prefilled; you see every line before anything is sent, and audio is never part of it.

Open it

Online: https://mdzahidh.github.io/rameau/

Offline: download index.html and open it —

open index.html

Any modern browser (Chrome, Safari, Firefox, Edge), straight off the filesystem. Nothing to install.

Use it

1. Drop in your recordings. One file onto guitar A, another onto guitar B — or press ● Record and play straight into the card. WAV, AIFF, FLAC, MP3 or M4A. The sample rate is read from the file's own bytes, so there is nothing to configure. One file on its own is fine — you get the full single-guitar analysis, minus the comparison views.
2. Name each guitar and set its kind (solidbody, hollow body, acoustic) on its card. The kind selects what is measured — a solidbody's signature is its pickup hump, a hollow body's is its air resonance.
3. Set the tuning you actually played — the same for both guitars. There is one tuning setting and the app assumes both guitars were tuned that way: the open-string axis, the anatomy regions, the string-stiffness reading and the guided take derive from it, and the sustain and dead-spot rows pair notes by pitch. Two guitars in different tunings are not a comparison this app makes.
4. Add a second take of each guitar. That is what tells the app how much a number moves on its own, and it starts to give verdicts from there.
5. Leave Level-match on while comparing. It cancels the broadband loudness gap between takes, so what you see is tone rather than output.

Then read down the page. Anything you don't recognise is clickable — labels, region names, the dots on the plots — and opens what it measures with your numbers already substituted into the formula.

How you record matters more than anything in this app. Change only the guitar: same player, same part, same tuning, same signal chain, same room, same gain staging. The in-app How to record guide covers it, and the guided take walks you through six steps.

Feedback and bug reports

Press ✎ Feedback in the app, or open an issue directly: https://github.com/mdzahidh/rameau/issues/new?template=feedback.yml

Tell us what happened or what you would like, in your own words. The form also asks — optionally, and it helps us know who the app is serving — whether you play professionally.

What it measures

• Long-term average spectrum — Welch, 8192-point Hann, 50 % overlap, resampled to a 700-point log grid from 60 Hz to 20 kHz. Octave smoothing is selectable (off, 1/12, 1/6, 1/3) and the active setting is always printed on the plot. A guitar with several takes is shown as the mean of its takes.
• Band energies over the region vocabulary you choose (band mix, EQ speak, guitar anatomy, solo EQ), with the boundaries drawn on the plot itself.
• Spectrograms — 2048-point Hann, 256 log-spaced cells, max-pooled per cell, on a shared colour scale so the two panes are directly comparable; zooming re-analyses the window at a finer resolution.
• Envelopes, aligned at each file's first detected onset.
• EQ match — a least-squares fit of standard analog band models against the smoothed difference curve, rendered on the face of a real device.
• Tone character — every row declares the evidence it needs and renders in one of four states (banded with a verdict, measured, partial, not measurable), with a reliability band measured from your own takes; harmonic numbers exist only for notes that pass a pitch comb check. The physics behind every row is in docs/THEORY.md.

Levels are dB relative to a full-scale sine throughout. Analysis parameters are printed in the page footer.

Repository

path	what it is
index.html	the entire application — the only shipped artifact
tests/verify.sh	the gate: five node suites, a headless-Chrome suite, tamper guards
tests/*.test.js	the suites; each extracts code from index.html and runs it under node
tests/audit_tone.js	the tone-panel audit on the three real takes in samples/
samples/	the demo pair and three real takes (Les Paul, SG, Majesty)
SPEC.md	the original commissioning prompt, verbatim, plus an append-only decision log
docs/ARCHITECTURE.md	DSP pipeline, parameter rationale, browser quirks, dead ends
docs/THEORY.md	the verified physics behind every measurement and every educational claim
docs/STORY.md	where the name and the app came from
docs/ROADMAP.md	what's being built next, in buildable pieces, and the rules for building them
.github/ISSUE_TEMPLATE/	the feedback form
LICENSE	MIT

Tests

./tests/verify.sh          # the whole gate, real Chrome included
./tests/verify.sh --node   # the node suites and tamper guards only, seconds
node tests/make_samples.js # regenerate the demo WAVs

No dependencies. Every suite reads index.html directly, so the tests always run against the shipped code rather than a copy of it.

Test hooks

Appending a query string drives the app for screenshots and manual checks:

hook	effect
?demo	load the synthesised demo pair (?demo=a / =b for one side)
?theme=bright|dark	force a theme
?open=all	unfold every collapsed panel — needed for full-page captures
?strings=1	show the open-string axis
?sgchord=E	overlay an open chord's harmonics on the spectrogram
?vocab=eq|anatomy|solo|mix|none	pick the region vocabulary
?scrollto=<card id>	scroll a card into view once the data has landed
?feedback	open the Feedback dialog
?debug	reveal the hidden demo-loading button and the ?load= hook

CLAUDE.md lists the rest, including the zoom, magnify and popover hooks.

License

MIT — see LICENSE.

The name

Jean-Philippe Rameau derived harmony from the overtone series in 1722, insisting that music theory was physics. His brother was named Claude — as is the AI this was built with. The About panel in the app tells the rest of it, including why a question about two guitars turned into this.