The Cost of Becoming Sound

"A system does not fail in one place. It fails a little, in order, at every point where a signal is asked to change its nature, and the only number that ever mattered is the sum."

A Number Leaves Home

Somewhere on a hard drive, or scattered across a network the way letters used to scatter across a sorting office, a piece of music exists as nothing more than numbers. Forty-four thousand one hundred of them a second, each one a tiny, exact statement: the air pressure at this instant was this value, and no other. There is no warmth in a number, no air, no sense of a room. There is only the value, sitting in silence, waiting to be asked for.

What happens next is a journey, and like most journeys, it is measured less by distance than by how many times the traveller has to change form to keep moving. Our number will become a voltage. The voltage will become a larger voltage. The larger voltage will become a current strong enough to move a physical object. The object will move the air in your room, and only then, fractions of a second after it left its quiet home on a hard drive, will it become the thing you actually came for: a sound.

Every one of those changes costs something. Not metaphorically. Literally, measurably, in volts and picoseconds and decibels. This is the story of that cost: where it is large, where it is trivial, and where the modern habit of inserting yet another box into the chain helps the traveller along, and where it simply gets in the way.

The Six Rooms It Has to Pass Through

Strip away the marketing and a digital playback system is six rooms with doors between them. A source, which retrieves the numbers. A converter, which turns them into a continuous voltage. A preamplifier, which decides how much of that voltage is allowed through. A power amplifier, which inflates it to something with real force behind it. A length of wire, which simply has to carry what it is given without opinion. And a loudspeaker, which performs the strangest trick of all: turning electricity into motion, and motion into sound.

That is the whole structure. Everything else that has appeared around it in the last decade, the switches, the isolators, the regenerators, the little boxes with their own small power supplies and their own confident names, lives inside the first room. They are furniture, not architecture. Whether a given system needs any of that furniture is a specific, answerable question, not a matter of taste, and we will get to it.

What does not change, no matter how the first room is decorated, is this: the signal changes medium three times on its way from number to sound. Digital to analogue. Small analogue to large analogue. Electrical to mechanical. Each transition is a place where something can be lost, and the loudspeaker, at the very end, performs by far the most violent of the three. Keep that fact in your pocket. We will need it later, because it changes where the money should go.

The Quiet Years Before Conversion

The first room, the source, has an easier job than people assume. Getting the numbers right is, in 2026, close to solved. Error correction and packet retransmission mean that what arrives at the converter is, with vanishingly rare exceptions, exactly what left the file. If your worry is "are the bits correct", you can largely stop worrying.

What you cannot stop worrying about, quite so easily, is what travels alongside those bits. A digital connection is never just data. It is also a small slice of the source device electrical life: the noise of its power supply, the wobble of its ground reference, the busy switching of its processor and memory. USB is the most exposed to this, because the specification itself ties the host five-volt rail and ground directly to whatever is plugged into it. None of that noise has to touch a single sample value to do damage. It only has to reach the converter clock recovery circuit, where it shows up not as an error in the data but as timing uncertainty, the audio engineer word for jitter.

Jitter does not change what a sample says. It changes exactly when that sample is allowed to become a voltage, and that small, repeated mistiming smears the fine edges of transients and slightly raises the noise floor around high-frequency content. It is not subtle in measurement. It is often very subtle in listening, because modern converters are built specifically to defend against it, a story told in full in The Digital Hierarchy. The short version: a converter that buffers incoming data and reclocks it against its own stable local oscillator has already done most of the necessary work before the signal reaches your ears. The quality of that internal clock, not the cable that fed it, is the deciding factor.

Grounding tells a parallel story, and it is told in full, with the necessary rigour, in Grounding and Shielding. What matters here is the shape of the problem: a noisy source shares its noise with everything downstream through a connection it cannot avoid making, and the fix is architectural (isolation, careful layout, a clean local supply), not cosmetic.

The Boxes Nobody Used to Need

Ten years ago, the first room of this chain contained a transport and a cable. Today it can contain a network switch, an Ethernet cable, a streamer, a USB cable, and a small isolating or regenerating device sitting just ahead of the converter, each sold with its own promise about what it will rescue. Before adding any of them, it is worth being precise about what each one actually does, because the precision is where the honest answer lives.

A network switch cannot fix timing. In a properly buffered streaming chain, audio crosses the network as ordinary data packets, asynchronously, and the streamer reconstructs playback timing from its own clock once those packets land. A switch has no more influence over when a sample becomes sound than the postal service has over what time you choose to open a letter. What a switch can do is electrical: a unit built around a cheap, noisy power supply can inject noise that rides the Ethernet cable into the streamer clocking and power circuitry, an effect that is real, measurable, and frequently smaller than the streamer own internal supply rejection. A switch with a quieter, linear supply addresses that noise at its source, and the general case for that swap, a dedicated linear supply in place of a stock switching wall adaptor, is made in full in Power Supplies; it applies to a switch exactly as it applies to a streamer or a DAC. The full mechanism behind switch-induced noise, and how to diagnose whether it is actually present in your own system, lives in Grounding and Shielding. The more complete answer removes the electrical path rather than quieting it, and it deserves its own explanation, because the phrase usually attached to it gets used loosely.

"Optical isolation" can mean two different things, and they are not the same fix. One is a small isolation barrier built into a device own network circuitry, a transformer or digital isolator sitting on the circuit board between the copper cable and the streamer internal electronics. It still uses an ordinary Ethernet cable for the actual connection, and the isolation, while real, has a finite rating, typically in the low kilovolts, and rejects some kinds of noise far better than others. The other meaning is literal: fiber optic cable, carrying the network connection as pulses of light rather than voltage on a wire. A media converter near the router turns the copper connection into light; fiber runs to the listening room; a second converter, or an SFP module plugged directly into a streamer built with the port for one, turns it back into an electrical signal only at the very last possible moment. Because light carries no current, there is no conductor at all between the noisy side of the network and the streamer for the length of that run, a stronger guarantee than any isolation transformer can offer.

It is not, however, a guarantee against everything. An SFP module still draws its own power from the streamer internal rail and performs the optical-to-electrical conversion on the streamer own ground, so the streamer internal design still has to be competent for the benefit to reach the output. And where the conversion happens in an external box rather than inside the streamer itself, that box needs a properly specified power supply of its own, the same linear-versus-switching question covered in Power Supplies, or it simply reintroduces the noise one hop later, closer to the equipment than it started. Fiber is the most complete way to remove a genuinely diagnosed network noise problem. It is also the most involved of everything in this section to set up properly, and exactly as unnecessary as the rest of it when no such problem has actually been found.

USB isolators solve a real and specific problem: the USB standard ties a computer ground directly to whatever it is plugged into, and a computer is a small factory of electrical noise. An isolator breaks that shared ground while still passing the data through. Whether adding one helps depends entirely on what the converter already does internally. A growing number of DACs already isolate their USB input on the circuit board, ahead of the conversion stage, and an external isolator placed in front of an already isolated input changes nothing measurable. Where the converter has no such isolation, an external unit can produce a genuine improvement. The distinction lives in the converter design, not in anyone impression of the sound.

USB regenerators are sold alongside isolators and often confused with them, but they answer a narrower question. A converter operating asynchronously, which is to say almost every modern DAC, governs its own sample timing from its own clock and merely reads data out of a buffer on its own schedule. A regenerator cannot improve timing the converter was never taking from the source in the first place. Where a regenerator earns its keep is power: in a converter that draws its analogue circuitry directly from the USB bus rather than from its own supply, a cleaner regulated five volts genuinely matters. In a self-powered, internally isolated converter, the same device is solving a problem that design does not have.

A ferrite core, clamped close to the device end of a USB or Ethernet cable, is the unglamorous exception to all of this. It attenuates broadband, high-frequency noise without touching the audio band, costs almost nothing, and has no downside. It will not fix a genuine fifty- or sixty-hertz ground loop, and it will not change jitter in an asynchronous interface, but as a first, cheap, reversible step before anything more expensive, it is the correct place to start.

The discipline that runs through everything in this section is the same discipline that runs through everything else in this article: identify the symptom, understand the mechanism, apply the narrowest fix. A system built around a converter with adequate internal isolation, a competent local supply, and a properly buffered asynchronous connection often needs none of this hardware at all, and that is not a controversial position. It follows directly from how these interfaces were designed to behave.

The Moment of Conversion

Everything before this point existed to deliver one thing cleanly to one place: accurate numbers, arriving with the least possible timing uncertainty and the least possible electrical noise, at the converter. Here, finally, the numbers stop being numbers.

A converter does three jobs in quick succession. It reconstructs a continuous signal from discrete samples, using a digital filter to remove the ultrasonic images that the sampling process creates. It turns that reconstructed signal into an analogue voltage or current. And it buffers that result so it can actually drive what comes next.

The filter choice, sharp or gentle, minimum-phase or linear-phase, is a genuine trade-off between time-domain behavior and frequency-domain attenuation, not a flaw to be solved. Any competently designed modern filter sits well past the point where that trade-off becomes audible under controlled conditions. The conversion step itself, in a well-built converter, achieves a total harmonic distortion plus noise figure of minus 100 decibels or better relative to full scale, a level no human ear has ever detected at normal listening volumes. The more practical danger is noise leaking in from the converter own digital processing core, and that is addressed with careful layout, isolated supplies, and balanced signal paths, the same toolkit described in Signal Integrity.

The output stage is often the part that lets a good chip down. It has to drive a cable and the input impedance of whatever follows while keeping its dynamic range and bandwidth intact, and a converter chip datasheet figures mean little if the surrounding circuit cannot live up to them. This is the point in the chain where the phrase "well engineered" stops being a slogan and starts being a measurable claim.

Two Gains, One Discipline

What leaves the converter is small. What needs to reach the loudspeaker is not. Between them sit two gain blocks, and both face the same essential discipline: add only what is needed, and add it cleanly.

The preamplifier task is attenuation, not amplification, most of the time: deciding precisely how much of an already-correct signal to let through. A poorly chosen volume control, a noisy potentiometer, mismatched channels, can introduce more thermal noise and channel imbalance than anything that happened upstream. The power amplifier task is the opposite and far more strenuous: turning a small, precise voltage into the current and force a loudspeaker actually needs. Distortion here clusters near clipping, but crossover artifacts in class-AB designs and supply-rail sag under heavy load both occur well inside the rated operating range of ordinary use.

The detail most often missed is gain structure itself, the relationship between the two blocks. A high-output source feeding a high-sensitivity amplifier forces the preamplifier to throw away decibels it never needed to generate, and every decibel discarded brings the system noise floor a little closer to the music. A well-matched chain keeps the signal as large as it can be at every stage without ever clipping, which is simply another way of saying: waste nothing.

There is a structural alternative worth naming here, because it changes this section shape rather than just its numbers. An active loudspeaker moves the power amplifier itself inside, or immediately beside, the cabinet, with a separate amplifier driving each individual unit directly. The two gain blocks above do not disappear, they relocate: the line-level signal leaving the preamplifier travels to the speaker largely unchanged, and the power amplification, one channel per driver, happens at the very end of the chain rather than several meters before it. What that relocation actually buys, and what it does not, is worth its own look once the loudspeaker itself is on the table.

The Wire That Carries It

A cable is, electrically, almost insultingly simple: a conductor, an insulator, sometimes a shield. Its behavior is fully described by four numbers, resistance, capacitance, inductance, and characteristic impedance, and all four are measured on ordinary instruments. Nothing about a cable requires faith.

For an analogue interconnect, capacitance and the source output impedance together form an unintentional low-pass filter. With a typical low-impedance solid-state source, a few hundred picofarads of cable capacitance places that filter corner far above 50 kilohertz, well clear of anything audible. With a high-impedance source, valve output stages and some passive preamplifiers among them, the same capacitance can matter, and the cable specification stops being a footnote. The full treatment of why geometry, not price, governs this behavior lives in Interconnect Geometry.

A phono cable is its own special case, because a moving-magnet cartridge behaves as an inductive source with a recommended capacitive load. Fifty picofarads of difference can move a resonant peak by several kilohertz, which is why the correct figure here is the one specified for the cartridge, not a vague reputation for "quality."

Digital cables are transmission lines, not wires in the ordinary sense, and the number that matters is characteristic impedance: 75 ohms for S/PDIF, 110 ohms for AES/EBU. Get it wrong and reflected energy returns to the source and corrupts the very timing the converter is trying to recover cleanly, a mechanism explored fully in Impedance Matching. Get it right, with a competently built cable, and the remaining differences between one correctly specified digital cable and another are electrically negligible.

Speaker cables live in the opposite regime: high current, very low impedance, where resistance is the parameter that decides everything. Enough resistance and the amplifier grip on the driver, its damping factor, weakens measurably. A cable with sufficiently low resistance and inductance for its length closes that question entirely. There is no remaining mechanism by which a more exotic cable would sound different, and Connectors covers the other half of that story, the part where contact quality, not conductor mythology, is what actually fails over time.

The Last Nonlinear Thing

Here is the fact that the whole chain has been quietly building toward. The loudspeaker is, by an enormous margin, the most nonlinear thing your music will encounter on its way to you. A genuinely excellent loudspeaker still distorts at one to three per cent at ordinary listening levels, climbing sharply from there. A competent converter, by comparison, distorts at less than 0.001 per cent. That is not a small gap. It is several orders of magnitude.

The loudspeaker earns that distortion honestly. The motor itself is nonlinear, its inductance and force shifting with the cone position. The suspension is nonlinear. The cone, at sufficiently high frequencies, stops moving as a single rigid piston and begins to flex and break up in ways no crossover can fully tame. The voice coil heats under sustained drive and its resistance climbs, quietly compressing the loudest passages. In a multi-driver design, the crossover itself adds phase shift, and the physical separation between drivers creates interference patterns that exist in the air, not in any circuit, and so cannot be corrected by any circuit either.

Some of that crossover-related loss, though, is not strictly unavoidable. A passive crossover sits directly in the high-current path between amplifier and driver, built from inductors, capacitors, and resistors that each carry their own series resistance, their own manufacturing tolerance, and, at real playback power, their own quiet nonlinearity. The network also dilutes the amplifier damping factor before it ever reaches the voice coil, because its components sit electrically between the two. An active loudspeaker removes that network from the power path entirely: the crossover is performed at line level, well before any amplification happens, and a dedicated amplifier then drives each individual unit directly, often from inside the cabinet itself. This changes nothing about the genuinely physical sources of distortion above it, cone breakup, suspension nonlinearity, and motor asymmetry exist regardless of where the amplifier lives, but it does remove the insertion loss, the component tolerance, and the diluted damping factor that a passive network adds on top of those physical limits. It is a real, measurable improvement confined to a specific part of the problem, not a cure for the whole of it. The one documented example of the approach in this knowledge base is Lipinski Sound, built around exactly this principle.

Then the room gets involved, a subject for its own telling, see Acoustical Basics, and reshapes what little of the original signal survived the journey this far. But even in an echo-free chamber, with the room contribution entirely removed, the loudspeaker alone imposes more amplitude error, more time-domain smear, and more nonlinear artefact than every electronic stage upstream of it combined.

This is not an argument against careful electronics. It is the argument for spending your attention correctly. The cleanest converter and the most disciplined amplifier, fed by a flawlessly isolated digital connection and terminated in a perfectly specified cable, still hand their signal to the one device in the chain guaranteed to alter it the most. The voltage reaching the voice coil can be, for all practical purposes, perfect. What leaves the cone already is not.

What the Descent Teaches

No stage in this chain can repair what an earlier stage damaged. A power amplifier cannot undo distortion the converter already added. A cable cannot restore bandwidth lost two rooms earlier. A USB isolator cannot fix a ground loop sitting in the analogue interconnects further down the line. And the most perfectly preserved electrical signal in the world, once it has passed through a loudspeaker distorting at one per cent with an uneven radiation pattern, is no longer perfect. It never had the chance to stay that way.

This is the cascade, and it dictates a genuinely practical order of operations.

• Get the source right. A clean, low-jitter digital feed into a well-designed converter is an easily achievable goal today, and for most converters with adequate internal isolation, no additional switch, isolator, or regenerator is required to reach it. Add one only where a specific problem has actually been diagnosed, never as a default.
• Preserve what the converter hands you. Competent amplification, correctly specified cabling, careful attention to grounding. Nothing exotic, just nothing wasted.
• Match the gain structure. Keep the signal as far above the noise floor as every stage allows.
• Spend where the damage actually happens. The loudspeaker and the room are where the largest, most measurable degradation in the entire chain occurs. A further improvement to an already-inaudible amplifier specification, or a third isolation device added to an already-isolated digital connection, buys nothing. A better-placed loudspeaker, or a treated reflection point, buys everything.

None of this is a sales position. It is simply what the cascade, followed honestly from one end to the other, requires. A number leaves a hard drive in silence, changes its nature four times, and arrives in your room as a sound. Most of what determines how well that journey goes was decided long before anyone reached for a cable, or a switch, or a regenerator: it was decided by where, along that journey, the real damage was allowed to happen, and where it was prevented.