Spokes.wiki Search About

Synthesis — Loudspeaker Design

The evolving thesis. Spun out of the hub _inbox loudspeaker-design cluster on 2026-06-28 from three sources arriving together: a T/S parameter explainer (ts-parameters-explained), an enclosure-design calculator (micka-enclosure-calculator), and a fundamentals primer (loudspeaker-primer-hackaday).

Current thesis

Loudspeaker design is predictive, not empirical: a driver's behavior in a box can be calculated in advance from a small set of measured numbers. The through-line of all three founding sources is the same loop:

  1. A moving-coil driver can't work in open air — its rear wave cancels its front wave (phase cancellation), so it needs a box.
  2. The Thiele/Small parameters capture the driver as a small-signal model (resonance Fs, damping Qts, air-compliance Vas, motor Bl, excursion Xmax, …) — themselves measured from a free-air impedance sweep (ts-parameters-wikipedia), and small-signal snapshots that drift with level/temperature/age and vary unit-to-unit, so the predictive loop is only as good as the measured inputs.
  3. Those numbers determine the enclosure — sealed vs. vented, the volume, the port tuning — and a calculator turns them into a predicted frequency/impedance/step/group-delay response before any wood is cut.

So the field's central move is replacing trial-and-error with a datasheet and a simulation. Wattage ratings say nothing about in-box performance; the T/S set is the real spec.

The design tensions (early)

The numbers have a standard, and the standard expired (research pass)

The predictive loop rests on measured inputs, and this cycle sourced the layer under them for the first time. A datasheet's T/S figures were produced to IEC 60268-5 — issued 1972, and in its 2003 form the document specifying these parameters, the anechoic chamber, the IEC baffle and the test boxes. It has been withdrawn, replaced by IEC 60268-21, IEC 60268-22 and IEC 63034 (iec-60268-5-replacement, geoff-hill, the spoke's first T1).

Why it broke is the part that matters here: the standard assumed a loudspeaker was "entirely passive." The active/DSP speaker this spoke reached from the hobbyist side (camilladsp, the DSP-correction thread) is exactly the device it could not describe. The DSP bridge is not a side-branch off the physics — it is what forced the measurement regime to be rewritten. The new set also adds chirp and multi-tone stimuli, 3D sound-field assessment, and Rub & Buzz defect testing, a fault class the old document had no way to name.

Practical residue for a builder: figures from the old and new regimes are not guaranteed comparable, and nothing here says how you tell which you are holding.

And the unit-to-unit caveat became a measured subject. The thesis above says T/S parameters are small-signal snapshots that vary unit to unit, so a careful builder measures their own driver — carried until now from ts-parameters-wikipedia and builder practice. ts-parameter-manufacturing-variation (AES Convention 125, 2008) measured twenty nominally identical drivers for exactly that spread, and its follow-on phase went further than tolerance: correlating parameter variation with failure mode and lifetime. The figures are behind the AES e-library and are not held here, so what this buys is the standing of the claim, not its magnitude.

Both sources arrived through the research pass against this spoke's first-T1 edge, which they close. The floor moves T1 0 → 2, and the shape of what is missing changes with it: the spoke now knows which standards its numbers come from and still holds none of their text.

The environment that cannot be designed

The thesis above works because the acoustic load is known before anything is built: a box you specify, a room that at least holds still. audio-dsp-hardbag-guide brings the first case in this corpus where that fails outright.

On a motorcycle there is no enclosure, and the noise floor is a function of road speed — in-vehicle-audio. Nothing in the predictive loop reaches that. You can still choose the driver by T/S parameters and predict what it does into a known load; you cannot predict wind at 100 km/h, and no calculator here models it.

The correction silicon got cheap, and nobody has measured what that costs. dspi (2026-10-07) puts the same five-function DSP correction camilladsp runs on a Raspberry Pi 5 onto a sub-$5 microcontroller instead. Two open projects now do crossovers, PEQ and time alignment at a two-orders-of-magnitude difference in silicon cost, and neither has a third-party measurement of audio quality against the other or against a passive network. The open question this spoke has carried since audio-dsp was written — what correction buys and what it costs versus a well- designed passive crossover — still has no instrument pointed at it, only two more implementations.

So audio-dsp arrives in a role the corpus has not had for it. loudspeaker-primer-hackaday calls DSP a shortcut, something that lets you correct rather than get the passive design right, and camilladsp is that shortcut as a tool: active crossovers and EQ replacing soldered networks. Here correction is not an alternative to design — it is the only remaining variable, because the environment half of the problem was never designable in the first place. That does not contradict the thesis; it bounds it. The predictive loop covers the transducer and its load, and stops at the door of the space.

One favourable asymmetry is worth recording, since it cuts the other way from everything else about a vehicle. Speaker placement is dictated by the chassis, so path lengths to the listener differ badly and time alignment becomes necessary rather than optional. But the listener does not move. Sweet-spot correction is a compromise in a room because someone will sit elsewhere; in a vehicle the seat is fixed, which makes per-position alignment worth more here than in any other case this spoke holds.

Evidential standing is poor and the page says so. The source is a store's blog recommending the category of product the store sells: T4, no author, no measurement, and its specific figures are manufacturer numbers passed through a reseller. One of them fails on arithmetic — 85 ms of delay is about 29 m of path difference, where a vehicle needs under 3 ms. The reasoning above rests on the geometry, not on the vendor. Car and motorcycle audio has its own competitive measurement culture and this spoke holds none of it, which is now a growth edge rather than a gap nobody had noticed.

Open questions

Measurements do predict preference, at about r = 0.70 (quality cycle, 2026-08-22)

The spoke has been running into the same wall from two directions. iec-60268-5-replacement has the standards body stating that its measurement regime "does not assess the perception and cognitive evaluation of the reproduced sound," and every calculator and alignment page here predicts a curve without ever saying which curve a person would rather hear.

olive-preference-model is the first source here that crosses the gap, and its numbers are more modest than the industry's use of them suggests. Four terms carry the 70-loudspeaker model: on-axis narrow-band deviation (31.5%), low-frequency extension (30.5%), in-room narrow-band deviation (20.5%), and in-room smoothness (17.5%). Flatness and bass, in other words, with directivity acting through the predicted in-room curve rather than on its own. The correlation is r = 0.70, so about half the variance in what listeners prefer is not explained by any of the seven curves.

Two things follow for this wiki. The sound-power result is a genuine negative finding worth keeping: a model built on 1/3-octave sound power correlated at r = 0.05 with preference, which is zero, and that was the measurement a consumer-testing organisation was publishing at the time. And the thiele-small-parameters design loop gains a target it never had — LFX, the −6 dB point, is one of the two heaviest terms, which is the alignment choice svs-sealed-vs-ported frames as a taste question showing up as a preference weight.

The caution stays on the page: this is Harman's own model, scored by Harman, and the peer-reviewed version is paywalled.

Growth edges

Ranked; each names the kind of source that would close it (see ../QUALITY.md → Growth edges).

  1. First T1 — CLOSED 2026-08-08 by the research pass: iec-60268-5-replacement (a standards-committee account of the measurement regime) and ts-parameter-manufacturing-variation (AES Convention 125). T1 0 → 2.
  2. The standards' own text. DROPPED 2026-08-30 at the curator's decision, after a third failed access route. Kept below as the record of why, because the failure is a fact about the field rather than about this wiki: the documents that define how every measurement in this spoke is produced are sold, not published. IEC sells them; AES sells them; the freely readable material is vendor marketing. A wiki built from open sources structurally cannot hold the specification its own numbers rest on, and that is worth stating plainly wherever those numbers appear. See iec-60268-5-replacement, which is as close as this spoke will get — a working-group chair describing the standards in a magazine. Original edge: The spoke now knows its numbers come from IEC 60268-21/-22 and 63034 and holds none of them. — needs: the standard itself, or a committee/working-group document specifying a measurement procedure · hunted 2026-08-30 — nothing ≥ bar; the re-specified asks are behind the same wall. The 2026-08-08 hunt failed on IEC's paywall and the edge was re-specified to accept an AES procedure paper or a working-group document instead. Both were tried today and both fail the same way: AES sells its standards too (aes.org routes every document through a standards store; the committee's own temporary PDF link had expired), and the AES standards blog entry for AES2-2012 is a 404. What is freely readable is vendor material — Klippel's slide decks on the IEC project, an Audio Precision application note, a consultancy's guide to reading a spinorama chart. All T3, and the Research Pass bar admits T3 only where the edge names a first-party fact, which "the procedure a standards body specifies" is not. So the edge is not scarce and not mis-specified; it is priced. Three access routes have now failed. The remaining honest routes are a library copy or buying one document, and neither is something a research pass can do on its own — it is a decision for the curator. Do not re-hunt on the same terms. (Earlier annotation: hunted 2026-08-08 — paywalled, nothing ingested.) IEC standards are sold, not published: iTeh, BSI, the ANSI webstore and GlobalSpec all return a catalogue page with the scope paragraph and none of the procedure. The scope alone is informative and is recorded here rather than paged — IEC 60268-21:2018 covers physical measurement of transfer behaviour between an arbitrary input and the acoustical output at any point in the near and far field, in both small and large signal domains, across transducers, TVs, portable audio, automotive and professional systems, and can account for the target application's acoustic boundary conditions. It explicitly "does not assess the perception and cognitive evaluation of the reproduced sound" — so the standard that replaced IEC 60268-5 draws the same physics/perception line this spoke keeps running into. No cooldown set: this failed on access, not on scarcity, and fourteen days changes nothing. The edge now needs a working-group document, an AES paper describing the procedure, or a library copy — not another search.
  3. No measured data, still. T/S parameters and alignments are explained and now documented as measured elsewhere; no measurement set lives here. ts-parameter-manufacturing-variation's figures are behind the AES e-library. — needs: an open driver test set, or a paper citing that study's numbers.
  4. Broaden toward general electroacoustics (microphones, amplifiers, room acoustics) rather than fragmenting. — needs: T1/T2 on any of the three.

Coverage edges (added 2026-08-08, at the curator's request for a wider backlog). These widen what the spoke covers instead of answering an open question above; one ordinary solid source closes any.

  1. How measurement is actually done. The spoke keeps failing to buy the standards' text and has no page on the instruments people measure with — REW, Klippel, a calibrated microphone and a gate. — needs: an AES paper on method, or the software's own documentation.

  2. The room. In-room response departs from anechoic by more than any driver choice, and room modes, boundary reinforcement and treatment appear nowhere. — needs: an acoustics text or an AES paper.

  3. Horns and waveguides. Directivity control is the third enclosure family beside loudspeaker-enclosure and passive-radiator, and it has no page. — needs: a design text or an AES paper with polar data.

  4. Listening tests as a discipline. HALF-CLOSED 2026-08-22. olive-preference-model (US 8,311,232 B2, T3) supplies the correlation the edge asked for — seven anechoic curves, six statistics, r = 0.70 across 70 loudspeakers, and r = 0.05 for sound power alone — but not from the venue the edge named. The AES papers behind it (6113 and 6190, 2004) are paywalled, so the wiki holds a commercially interested party's own account of how well its own model works. Successor, new 2026-08-22. HUNTED 2026-08-28 — half-closed, and the residue is sharper. spinorama (T1) is the score computed outside Harman on 1,000+ speakers Harman did not test, in open source, from third-party measurements. It is not a re-fit: Olive's coefficients are used unchanged, so nobody has independently re-derived the model and that half stands open. What the hunt did close is a different question nobody had asked — the score moves with the measurement rig, by 0.5–1.5 points for smoothed data against a 0.6-point significance threshold, per the maintainer's own documentation. — needs still: an independent re-fit of the regression, or a listening-test validation on speakers Harman did not test, T1/T2. Not eligible again before 2026-09-11. Second successor: the model is silent on how the listening tests themselves were run (double-blind, positional substitution, training of listeners), which is the discipline half of the original edge — needs: a published listening-test methodology, T1.

  5. Vehicle audio, measured. New 2026-08-10. in-vehicle-audio entered on a single T4 retailer page plus geometry, and car audio has a large competitive measurement culture this spoke holds none of. — needs: an in-car response measurement or an AES paper on automotive acoustics, not another installer guide.

  6. What DSP correction costs. New 2026-08-10. audio-dsp is now paged as the correction layer and the corpus cannot say what it gives up: fixed vs floating point, cumulative filter error, or the excursion ceiling that bounds any EQ boost. — needs: a DSP text or an AES paper; ../music-tech-wiki's digital-filter-audibility is the nearest thing anywhere in the hub.

Bass costs time, and the market doesn't pay for either (research pass, 2026-08-10)

A directed research pass on hi-fi bass reproduction — no source handed in, two fetched — added the timing axis the spoke had been gesturing at and one adjacency it had never touched.

The bass bargain, stated. Every mechanism that buys low-frequency extension or output is resonant, and resonance returns energy late. Ports, passive radiators, high-pass filters and EQ boosts all pay in group-delay, and milliseconds cost more at 30 Hz than anywhere else because a cycle is 33 ms long. This sharpens the founding sealed-vs-vented tension from a preference into a currency: extension and output are bought with time, and the tuning frequency sets the price.

The reframe comes from svs-sealed-vs-ported (T3, and a port vendor): the timing penalty tracks where a box is tuned, not the presence of a port — tune below 20 Hz and the phase rotation falls under the music band. Unverified, and exactly the claim that seller benefits from, but it is a testable statement rather than a slogan, and it recasts "sealed for music" as a shorthand for "most ported subs are tuned too high."

room-gain is the second place the loop stops. The spoke already knew the room isn't designable (in-vehicle-audio was the extreme case). Room gain names the specific bass consequence: below the lowest wavelength that fits, the speaker pressurizes a volume and output rises, so the simulated curve and the heard curve diverge exactly where the design work was hardest. A sealed box's gentle fall can be spent against that rise — a designer choosing an alignment for a room they will never measure.

And then the market ignores all of it. hifi-value-guide-potteries prices used hi-fi by brand, condition, completeness and scarcity. Not one measured quantity appears — no T/S figure, no response curve, no group-delay plot. Speakers span £30 to £3,000+ a pair on badge and box condition. Recorded as loudspeaker-resale-value, an adjacency (economics, not electroacoustics) that is kept because it is the only place a design decision gets priced by someone other than the designer. It doesn't contradict the predictive thesis — it bounds its reach: the loop predicts quality, and quality is not what clears the used market.

One live sub-thread: passive boxes hold value because they cannot obsolete, while audio-dsp-carrying active speakers have an end-of-support date. The spoke's DSP thread therefore has a resale cost attached to it that nothing in the corpus has measured.

Growth edges added: #11 a neutral, measured sealed-vs-ported group-delay comparison (the figures in circulation are content-farm numbers; Audioholics' myth piece 403'd); #12 does measured performance predict resale value, controlling for brand — answerable with a used-market scrape plus a measurement database; the Headphonesty depreciation dataset (Feb 2026) also 403'd and should be retried.

The score is now computed by outsiders, and the measurement rig moves it (research pass, 2026-08-28)

olive-preference-model left the spoke holding a preference model fitted, scored and reported by the company selling the loudspeakers. The research pass went looking for somebody else's fit of it and did not find one. What it found instead is spinorama (pierre-aubert, GPL-3.0, T1): the same regression, computed by a party with nothing to sell, over 1,000+ speakers Harman never tested, from measurements made by third parties — Audio Science Review, Erin's Audio Corner, Princeton 3D3A, Audioholics and manufacturers.

That is application at scale, not validation. The coefficients are still Olive's and nobody has re-derived them, so the edge survives in its narrow form.

The new finding is about the instrument rather than the model. The site's own documentation warns that "smoother or less precise data can yield a higher score than an Klippel generated set" and recommends subtracting 0.5 to 1.5 for highly smoothed data — against a stated significance threshold of 0.6 between two scores. The correction is one to two and a half times the smallest difference the score claims to resolve. So a published preference score is comparable only between speakers measured on comparable rigs, and nothing in the number says which rig it came from.

This is the same shape as the standards problem the spoke keeps meeting from the other side. iec-60268-5-replacement has the committees rewriting the measurement regime because the old one was built for a passive device; here the regime is fine and the execution varies enough to swamp the metric. Both are the gap between a number and the conditions that produced it, and this wiki's whole predictive loop — datasheet to simulated box to response — runs on numbers whose conditions are usually unstated.

Two limits worth carrying to any page that quotes a score: it says nothing about maximum output or distortion ("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, so the site refuses to print a number at all for surround, in-wall and column speakers.

Contradictions / tensions

Cross-spoke adjacency

Index — Loudspeaker Design Wiki

Catalog of every page, grouped by schema.org @type. Spine: synthesis (thesis), log.md (history), this file (catalog). Spun out of the hub _inbox loudspeaker-design cluster on 2026-06-28 (3 sources).

DefinedTerm (concepts / components)

Product (loudspeakers)

TechArticle / WebApplication (sources)

Person / Organization (entities)

Synthesis

Notes

Related pages

Links to

Linked from