How it works

The signal chain

A and B are two complete systems running side by side, each a chain of stages the music passes through.

1. Source

   Digital plays the recording as it is. Vinyl adds what a record and turntable do:

   • Bass narrowed toward mono when the master disc (the lacquer) is cut
   • Treble held within what a record can carry: cutting systems limit how fast the cutting stylus accelerates. Ours turns the band above 4 kHz down, briefly, whenever it would pass 830 g, the setting of Neumann's 1974 acceleration limiter; real records measured by Jim Lesurf (Hi-Fi News, 2007) stay within about 1,000 g. Very bright recordings lose a little top treble, as they do on vinyl; most of ours are untouched. The release time (50 ms) is our choice
   • The cartridge's frequency response (how evenly it plays each pitch) and channel crosstalk (left and right leaking into each other)
   • Tracing distortion (the stylus tip can't follow the groove's exact shape), made in the groove and calibrated to Shure's measurements by stylus shape, then surface noise, clicks and scratches
   • Treble lost near the label, where the groove is slower: Miller's model (as restated by CBS Laboratories, 1962) with each cartridge's tracking force. Conical tips use their own radius; for other shapes the radius is fitted to Hi-Fi World's measured inner-groove losses (our choice)
   • Mistracking: on the loudest passages the stylus leaves the groove wall. Each cartridge's limit is its measured tracking ability (µm of groove swing at 300 Hz; maker's or typical figures where none was published), checked against Hi-Fi News's test-record results; the limit at higher frequencies scales from Shure's V15 IV measurements (our choice). Most music never reaches it
   • Wow and flutter (speed wobble)
   • Rumble (low-pitched turntable noise) and record warp, both shaped by the tonearm resonance. Each cartridge plays on an arm that suits it: the turntable's own arm when the resonance lands in the usual 8–12 Hz window, otherwise the effective mass that aims for 10 Hz, kept within our measured arms' 7–15 g (our choice, as an owner would change the arm or add headshell weight). You can switch to the stock arm to hear the mismatch.
   • Distortion as listeners describe it: the Why panel's Distortion section plays test tones (100 Hz and 1 kHz, at the music's loud peaks, about 12 dB above its average, our choice) through the source and amp (and the speakers too when they differ; identical speakers would add the same distortion to both and hide the difference), and shows the harmonics each system adds, plus two standard intermodulation tests (SMPTE: 60 Hz with 7 kHz; twin-tone: 19 with 20 kHz). These describe the kind of distortion; the expected difference rates how audible it is on the music
   • The phono stage, from published bench tests: its RIAA error (how far it strays from the curve that undoes the record's treble boost and bass cut), its hiss, its input overload, and its 2nd harmonic where it matters (the EMT 128: 0.3% at its nominal input, growing with level). Hiss and headroom are set against each cartridge's output, through the moving-coil input for cartridges under 1.5 mV and the moving-magnet input otherwise (our choice). Only stages with that input are offered, as an owner would choose: the Fosi has no moving-coil input and the EMT 128 no moving-magnet one. The hiss is white (our choice, close to the measured spectra). Overload clips the cartridge's signal, with the clicks and surface noise, at the measured point at 1 kHz; where headroom falls toward 20 kHz, a shelf above 5 kHz (our choice) lowers it to the 20 kHz figure. The Pro-Ject RS2 is the default; a stage's loading of the cartridge and step-up transformers are not modeled yet.

   The MP3 comparisons and the MP3 source in Build your own use one of our recordings, "When I Cross The River", encoded ahead of time so it lines up with the original sample for sample.

2. Amplifier

   Calibrated to each amp's bench measurements:

   • 2nd and 3rd harmonic distortion (added overtones), falling with level as in real amps
   • Clipping (flattened peaks when the amp runs out of power) at the measured power, soft or hard, and lopsided where the real amp clips that way
   • Hum and hiss at a fixed voltage, so they're audible through very efficient speakers (ones that play loud from little power)
   • The amp's output impedance (its internal resistance) against the speaker's measured impedance (its electrical load at each pitch). With tube amps, the speaker reshapes the tone and distortion rises where its impedance dips.

3. Speakers

   • Measured frequency response straight ahead (on-axis) and at angles (off-axis), smoothed to 1/6 octave
   • How the sound spreads: off-axis sound drives the room's reflections and reverberation
   • Loud-level behavior in four frequency bands: power compression (output falls short as the voice coils heat up), distortion that grows with level, and the limiter of active speakers (ones with built-in amps)
   • Crossover timing: how far apart in time the sound from each driver (woofer, tweeter) arrives

   Some comparisons edit a speaker (a bump, dip or bass roll-off, or another speaker's off-axis sound); the card says which.

4. Room

   • Speakers and listener in an equilateral triangle, speakers angled in toward you (toe-in)
   • Direct sound and early reflections, each shaped by how the speaker spreads its sound at that angle (its dispersion) and by the wall it bounces off.
   • Bass below about 200 Hz, where rooms boom: each resonance dies away at its own rate, set by the walls it bounces between (Kuttruff, Room Acoustics). The absorption data start at 125 Hz, so real rooms with thin absorbers may ring longer in the deepest bass.
   • Reverberation (the sound that lingers after each note) for each octave band, set by the total sound the speaker sends into the room
   • Air absorption, and surface materials from a published absorption database

   Some comparisons, and the Room setup and EQ tiles in Build your own, move the speakers or seat, or add subwoofers, room correction or tone controls. Subwoofers take over below 80 Hz through a steep (4th-order) crossover. Each is sealed, −3 dB at 20 Hz (our choice of a typical one), and set at the main seat to be as loud from 40 to 80 Hz as a speaker in the main speakers' places. Room correction is modeled on how Dirac Live is commonly used: it evens out the bass toward the neutral target, below 300 Hz as many owners of good speakers set it, and not below where the system's bass runs out (the main speakers' −6 dB point, or 25 Hz with subwoofers). It cuts peaks and boosts dips by up to 6 dB, but leaves dips deeper than 12 dB alone: they come from sound cancelling at the seat, which boosting can't fill. Those limits are our choices. It models only level, not Dirac's timing correction, and lowers bass peaks without much shortening their ringing, which only absorption does. The tone controls are the taste test's bass, presence, treble and air filters. They act after the amp and speakers, so a bass boost doesn't make them work harder, as it would in a real system played loud. "Use my taste result" sets them from your latest taste test.

   The test-room standard. When audio researchers run listening tests to find small differences, they use rooms built to an international standard, ITU-R BS.1116. Such a room is mid-sized (a floor of 20 to 60 m²), its proportions fall within the standard's limits, and sound dies away quickly and evenly. In our treated room (5 × 4 m, 2.6 m high) the standard asks for a decay of about 0.2 s through the midrange, give or take 0.05 s, and up to 0.3 s longer in the deepest bass. Reflections arriving in the first 15 ms must be at least 10 dB quieter than the sound straight from the speakers. The point is a room that adds little of its own, so differences in the equipment are easier to hear. That makes it a reference, not the only good sound: many people prefer a livelier room at home. On the site, the treated listening room is built to this standard, and the Room sound section compares any room with it.

   The Room sound section of "What's different, and why" measures the simulated room as an acoustician would a real one. Decay time uses the ISO 3382 method, against the target in ITU-R BS.1116, the international standard for rooms used in listening tests (the "test-room standard" on the site). Decay and consistency are measured at nine spots: the three seats of a sofa (0.6 m apart), and 0.3 m in front of and behind each, our choice of a typical listening area. Early reflections list the strongest in the first 20 ms and check the BS.1116 rule that those in the first 15 ms be at least 10 dB below the direct sound from 1 to 8 kHz.

5. Your ears

   A simple model treats your head as a sphere and places each sound by its direction, adding the small delay between your ears and the shadow your head casts on the far ear. It is generic, not fitted to your head, so some people hear the speakers closer or inside the head.

Fair comparisons

• Equal loudness. A louder system almost always sounds better, so we match both at the end of the chain with broadcasting's standard loudness measure (ITU-R BS.1770), measured on a second pass so the speakers' voice coils are warm, as when you listen.
• Instant switching. Both systems play all the time; A and B only choose which you hear, so no gap gives the switch away.
• Music you know. Differences are easier to judge in songs you know, so the questions invite you to upload some, with our recordings as the backup. A few comparisons need their own recording, for a test tone or very deep bass.
• Original recording. This button plays the music with no system or room, at the same loudness.
• Questions about what you hear. Where a comparison has a clear direction, it asks about the sound ("Which one has deeper bass?"), not which system is which.
• Blind by default. In the questions, which system is A, which one the question asks about, and the order of names are random each visit.
• Blind test. In Build your own, "Blind test this comparison" plays X, secretly A or B, for 10 or 16 rounds; the result shows how often guessing would do as well.
• Playing level is separate from your volume. "How loud will it be played?" sets how hard the system works (70–105 dB at your seat), bringing out compression, clipping and distortion. Your device volume sets how loud you hear it.

Headphones or speakers

Use the toggle at the top to say what you're listening on. On headphones (recommended), the site simulates the whole system, with two virtual speakers in the chosen room. On speakers, your own speakers and room play the result, so the site applies only the amplifier and the tonal difference between the chosen speaker and a reference (KEF LS50 Meta). Room choices are off, so the questions about rooms, placement, room correction, subwoofers and off-axis sound need headphones.

One speaker or two. On headphones, Build your own can play one speaker in the center instead of a pair: in loudspeaker researcher Floyd Toole's tests, listeners heard speaker differences more easily with one than with two.

The expected difference

The expected difference ("Likely inaudible", "Subtle", "Clear", "Obvious") is our estimate of how easy the gap is to hear: the highest rating any of six measures gives. It can't judge MP3 compression, which hides its errors under louder sounds.

Space steps. A step is the smallest change listeners noticed in published tests that switch between two sound fields, as an A/B comparison does:

• Decay time: 5% (ISO 3382-1, after Seraphim's 1958 tests), but at least 0.04 s, the threshold Niaounakis and Davies measured in small listening rooms with short decays.
• Direct sound against the reflections: about 2.4 dB when they are about equal, rising to about 3 dB at 10 dB and about 6 dB at −10 dB (Larsen and colleagues, 2008).
• Apparent width: 0.065 in 1 − IACC_E3, the standard measure of how wide a source sounds: one minus how alike the two ears' early sound is in the first 80 ms (Okano, 2002). It is computed for each speaker at your ears, from 500 Hz to 2 kHz.

Image steps. BBC researchers (Harwood and Shorter, 1964) played a sound from one speaker, fed a little of it into the other, and asked listeners where the edge of the sound had moved. On average they noticed crosstalk about 19 dB below the music when it was the same at every pitch (speech), and about 13 dB below at 10 kHz when it rose toward the treble (percussion). A step is the shift they just noticed; we take a cartridge's separation at 1 kHz against the first test and at 10 kHz against the second. Every magnetic cartridge we model stays under one step. The 1960s ceramic, with 7 dB of separation at 10 kHz, moves the treble of a sound at the side about 8°, roughly four and a half steps. A sound panned fully to one side is the most sensitive case, so most of a mix moves less.

Thresholds vary a lot between listeners and studies: later tests put the decay threshold anywhere from about 3% to over 20%, and a 2023 study found it depends on each listener's hearing and experience. Our simple ear model (a sphere without outer ears) makes the width figure approximate. How steps map to Subtle, Clear and Obvious is our choice.

Distortion and masking. Loud music hides distortion near its own notes; the same distortion in the top octave, or in a quiet passage, is easier to hear, and hearing itself fades above about 15 kHz. So distortion is rated with the ear model of ITU-R BS.1387 (PEAQ), the standard for measuring audible impairment, as described by Kabal (2002): the outer and middle ear's response, the ear's internal noise, masking spread across neighbouring pitches and forward in time. The standard counts a moment as disturbed when, in any band, the distortion rises more than 1.5 dB above the masked threshold; we call distortion likely inaudible when no moment is disturbed, since a single disturbed moment (clipping on one drum hit, say) can be heard. No published scale grades distortion above that threshold, so Subtle, Clear and Obvious use 10 dB steps, as the earlier rule did (our choice). The music is taken to play at your chosen listening level. Background noise (hiss, hum, clicks) is rated separately, in quiet moments.

The other bands are our rules of thumb, not listening-test results; the anonymous tally of your guesses will show how well they hold up.

Live curves in the Why panel show what each system and its room do to the music at your seat, over the last second. Each compares the sound at your seat with the original recording, lined up on the midrange, so the music's own shape cancels out. Where the curves part, A and B sound different. A faint dashed line, also on the seat chart, marks a neutral target: flat below about 300 Hz (the recording's own bass, which room correction and subwoofers aim for), then sloping gently down, as a neutral speaker does at a seat, because rooms and ears soften the treble. Listeners prefer that slope to a curve forced flat. It comes from the Revel F328Be simulated at the same seat, in the room both systems use (the treated room when they differ).

Find your taste

Each round starts from a neutral reference and changes one thing: bass, bass tightness, treble, presence (upper midrange), air (the top octave), room sound, how widely the speakers spread their sound, and optionally tubes, vinyl and punch at loud levels.

• Three lengths. Quick taste is 7 rounds (two votes each on bass, treble and room sound, plus one identical pair) and gives leanings, not a tested profile. You can carry on into the full test, which Sound only and Sound and gear run.
• Your own music. Each song in your playlist plays in the rounds it suits, and our recordings fill the rest; ★ in the music menu marks the one we chose.
• Saved as you go. Answers are kept in this browser after every round, so you can Save and exit and resume later.
• Round by round. The results show your votes for each trait and how often random picks would split that way.
• Listen longer. An optional six-minute step plays your sound and its opposite, three minutes each without switching, then asks which you'd rather keep.

To keep the result honest:

• Sides are random. With our recordings, each trait is heard on different music; with your playlist, your songs take turns.
• Some pairs repeat with the sides swapped, to measure consistency; a few are identical, to catch hearing differences that aren't there.
• A trait counts as a preference only when your votes pass a statistical test; traits that lean one way without passing get extra rounds.
• The room rounds change only reverberation and reflections. Livelier rooms also change tone (about 7 dB more bass around 63 Hz and up to 3.4 dB less treble at your seat than the treated room), so these rounds match the rooms' bass with room correction (within 2 dB, though below about 40 Hz the livelier rooms stay lower) and their treble with a gentle lift (within 0.1 dB). That way, liking more bass or less treble isn't counted as liking a bigger room.
• A final check plays your predicted favorite against alternatives.

Recommendations match your profile against traits computed from each speaker's and room's measurements, never hand-labeled. The same traits label the options in Build your own, relative to the others here, not to an absolute standard.

How far to trust it

Every part carries a label saying where its numbers come from:

Presets closely approximate the named products but aren't exact replicas. Also:

• The head model is generic (see Your ears).
• Speaker measurements are smoothed to 1/6 octave, which hides very narrow features.
• We don't model how each turntable drive type (belt, idler, direct) holds speed against the stylus's drag: we found no published measurement to base it on.

Sources

What each part is based on

Every cartridge, turntable and record number with its source page: data/vinyl/README.md.

Room materials

Absorption coefficients per octave band (0 = fully reflective, 1 = fully absorbing).

Speaker measurement licenses

Most speaker data come from Audio Science Review and Erin's Audio Corner, measured with a Klippel near-field scanner: a robotic rig that measures a speaker up close from many angles and removes the room's effect. The Princeton 3D3A Lab measured the Quad ESL-57 and Avantgarde Solo in an anechoic (echo-free) chamber. Audio Science Review LLC holds the copyright in its data, which we use under CC BY-NC-SA 4.0 (non-commercial). We smoothed it to 1/6 octave, converted it to relative responses, and share the result under the same license. The Erin's Audio Corner data (10 speakers, plus most of the distortion and compression figures) and the Princeton 3D3A data (2 speakers) have no published license; permission is pending.

References

Privacy

Your own music never leaves your browser. "Keep my songs in this browser" stores the files here until you choose "Forget my songs". Your guesses, blind-test scores and taste results, including Listen longer, go into an anonymous tally of answers, systems compared and listening conditions, with no personal data; your songs appear only as "own", or, in the blind test, as one of our recommended songs when your file matches one. Your progress and latest taste result are kept in this browser.