Spinorama (spinorama.org)
pierre-aubert's open database of loudspeaker measurements — 1,000+ speakers, each rendered as the CTA-2034 spinorama curve set with the preference score computed from it. GPL-3.0, Python / Cython / Rust, 165★ / 15 forks, repository created 2020-02-08 and pushed 2026-08-27. Read from the site documentation, the repository README and the GitHub API on 2026-08-28. Found by the research pass hunting growth edge 8's successor: somebody else's fit of the preference model.
Term versus project. "Spinorama" names the CTA-2034 curve family — on-axis, listening window, early reflections, sound power and their directivity indices. This page is the project that publishes them at scale; the underlying measurement standard is the same one iec-60268-5-replacement describes the committees rewriting.
What it actually is, and what it is not
It computes Olive's preference score on speakers Harman did not test, using measurements it did not make. The measurement sources it credits are third-party reviewers and labs — Audio Science Review, Erin's Audio Corner, Princeton 3D3A, Audioholics, individual reviewers, and manufacturers' own data.
So it is an independent implementation applied at scale, not an independent validation. The regression coefficients are still Olive's, fitted on Harman's 70 loudspeakers in Harman's room. Nobody here has re-derived them, and the edge that sent the research pass out stays open in that narrower form. What changed is that the score is now computed by a party with no speaker to sell, in open source anyone can read, over a population two orders larger than the one it was fitted on.
The caveat that matters most, stated by the person computing the scores
"Be also mindfull than smoother or less precise data can yield a higher score than an Klippel generated set of data for example. You can easily remove 0.5 to 1.5 to the score for highly smoothed data."
Measurement quality biases the score upward. A speaker measured on a Klippel Near Field Scanner is scored on data that resolves narrow-band deviation the score is built to punish; a smoothed curve from a lesser rig hides exactly the wiggles that cost points. The recommended correction — subtract 0.5 to 1.5 — is between one and two and a half times the site's own significance threshold:
"Note that a difference between two scores is only significant if the difference is greater than 0.6."
Put those two sentences together and a published preference score is comparable only between speakers measured on comparable rigs. That is a limit on every score table on the internet, this one included, and it is not visible in the number.
Two further limits the documentation states outright: the score "does not tell about maximum output
or about distorsion" — "you can get the same score for a portable bluetooth speaker and a large
tower" — and the predicted in-room curve assumes a rectangular room with standard reflections, so
the site prints *** instead of a number for surround, in-wall and column speakers, and warns that
"all bets are off" for omnidirectional designs.
Where it fits this spoke's argument
olive-preference-model established that flatness and bass extension carry most of the preference variance, at r = 0.70 — half the variance unexplained. This page adds the operational half: even the explained part is only as good as the rig that measured it, and the corpus can now say what the number is silent about. The thiele-small-parameters design loop predicts a curve; the score grades the curve; and the grade moves with the measurement gear before it moves with the speaker.
The LFX term is reproduced in the documentation as Olive defines it — the log₁₀ of the first frequency below 300 Hz where the sound power curve sits 6 dB under the listening-window mean from 300 Hz–10 kHz, log-transformed to linearise it against preference, and taken from sound power rather than on-axis "because it better defines the true bass output of the loudspeaker, particularly speakers that have rear-firing ports." That is a more precise statement of the bass-extension term than this wiki held, and it is not the −3 dB point builders usually quote.
What it cannot settle
No listening tests. No re-fit. No error bars on the underlying measurements, which come from a dozen rigs of different quality — the site's own warning is the acknowledgement of that. And the aggregation is a convenience sample: speakers get measured because a reviewer bought or was sent one, so the 1,000+ is not a sample of the market in any statistical sense.
Related
olive-preference-model · pierre-aubert · sean-olive · thiele-small-parameters · iec-60268-5-replacement · loudspeaker-resale-value · synthesis