Total Harmonic Distortion (THD) Explained

DAC and amplifier boxes now routinely show a THD+N figure with more zeros after the decimal point than the last model, and it reads like an obvious ranking: more zeros, cleaner sound, better product. What the box almost never tells you is the volume level, pitch, load and bandwidth the figure was measured at. Without those four, a THD or THD+N percentage isn't just incomplete. You can't compare it with any other maker's number, and sometimes not even with the same maker's number for the same product measured a different way.

THD and THD+N measure two different things

Total harmonic distortion, or THD, is a ratio: how strong the extra overtones a device adds to a pure test tone are, compared with the tone itself. Test-equipment maker Audio Precision's glossary explainer on the two measurements says plainly that "noise is not included in a THD measurement". A modern analyser picks out only the exact pitches where the test tone's overtones fall and adds those up, a result the same page describes as "not influenced by DUT noise or spurious interfering signals."

THD+N is the one you'll usually see published, and it works differently. Instead of picking out overtones, it filters out the test tone and adds up everything left over (overtones, hiss, hum, anything else) into one number. AES17, the Audio Engineering Society's own standard for measuring digital audio equipment (read here in a copy hosted by l7audiolab.com), defines it in exactly that spirit: "harmonic distortion and noise is the ratio of the output noise and distortion level to the output signal level. Both levels shall include all harmonic, inharmonic and noise components." The same clause of AES17 immediately warns what the number is and isn't good for: "the results are indicative of anomalies in device behavior but may not be indicative of audible performance." A single THD+N percentage is a health check for the device, not a prediction of what you'll hear.

The percentage means nothing without level, pitch, load and bandwidth

AES17 doesn't let one THD+N reading describe a device. Its procedure calls for testing "at –1,0 dB FS and repeated with a sine wave at –20 dB FS," then separately "at each level from 0 dB FS to –80 dB FS in steps no larger than 10 dB," and again "at each octave frequency from 20 Hz to one-half the upper band-edge frequency". The same circuit's distortion really does change with level and pitch, so a single spot check would give the wrong picture either way. Audio Precision's own comparison of a complete and an incomplete THD+N specification makes the same point about what a spec sheet owes you: a useful figure reads "THD+N less than 0.01%, 1 Vrms, 20 Hz - 20 kHz, unity gain, 20 kHz BW," while "THD less than 0.01%" on its own tells a buyer almost nothing.

Bandwidth matters for a reason that's easy to miss: it can quietly leave out the very overtones being measured. An independent write-up on audio testing at nihtila.com works through the maths: "if the input bandwidth is, say 22 kHz, the third harmonic of the signal will only be included up to fundamental frequencies of about 7 kHz". So a fixed measurement bandwidth quietly limits which overtones of a high test tone get counted at all, and a maker choosing that bandwidth is also choosing, on purpose or not, how much of the real distortion gets reported.

Level and load show up plainly on real spec sheets, not just in theory. Yamaha's own technical data sheet for the PX3 power amplifier, hosted by the dealer saleswl.com, lists "Total Harmonic Distortion 0.1% (1kHz, 10W), 0.3% (1kHz, Half power)". Same amplifier, same pitch, and the published figure triples just from turning the output up toward its maximum. The sheet rates the PX3 into 8 Ω, 4 Ω and 2 Ω speakers at different power levels, but doesn't say which of those either distortion reading was measured into. Another maker's convention looks nothing like that: Crown's own product page for the XLC 21300 gives a single figure across the whole range, "THD: <0.5%, 20 Hz–20 kHz," with no power level or load at all. Neither is dishonest, and you can't line one up against the other. One is a worst case across the whole audio range; the other is two best-case snapshots at a single pitch; and neither page says what speaker load the amplifier was driving when the number was taken.

Why makers' figures don't line up

Nothing forces a maker to publish a THD or THD+N figure the way AES17 measures it. The standard only asks that a report following it say so: it recommends a note saying results were "measured in accordance with AES17", which most consumer spec sheets don't include, so there's no guarantee they used the same level sweep, pitch sweep or bandwidth the standard requires. Amplifiers usually follow a different standard: AES17 itself lists IEC 60268-3, the sound-system-equipment standard covering amplifiers, among its own references, and that's as much as we can responsibly say about it. We name it only because AES17 does, since IEC 60268-3 itself sits behind a paywall this article did not access, so we've left out what it requires rather than guess. Either way, the practical upshot is the same: a THD figure with no level, pitch, load or bandwidth given tells you nothing compared with a rival's figure, except that the conditions might, or might not, have been the same.

One number hides which overtones you'd actually notice

THD and THD+N both squash every overtone into one percentage, but our ears don't treat all overtones the same. Dolby engineer Louis Fielder's AES convention paper on perceptual assessment of headphone distortion proposes judging distortion with "a critical-band spectral comparison of the distortion and noise to an appropriate masked threshold", with separate masking curves for different pitches rather than one blanket limit. That's because whether you hear an overtone depends on where it lands compared with the original note, and how much nearby sound hides it, not just on how big it is. Two devices with the same total THD, one mostly a well-hidden high overtone and the other mostly an exposed low one, wouldn't sound equally clean, and one percentage can't tell them apart. By its own summary, that model was built for a pure tone played through headphones, with only a brief mention of speakers. It isn't a claim about music, and we haven't tried to find a source for that.

That's also the honest answer on audibility. There's no single percentage we can cite below which distortion "becomes inaudible" for every kind of overtone and every kind of music. Fielder's paper builds a multi-band model instead of offering one number precisely because a fixed limit falls apart when you look at how differently a 2nd and a 9th overtone, or a pure tone and a piece of music, get hidden. Where we can't cite a study we've read in full for a specific number, we say so, instead of repeating one of the many unsourced percentages in buying guides and forum threads.

Distortion is outside what our tools measure. There's no THD meter here, and nothing about a browser's audio output makes it possible to build one reliably. What a demanding pair of headphones does to volume and frequency response, rather than distortion, is covered in headphone impedance and why it matters. How cleanly digital audio stores a waveform in the first place, a separate question from how a real amplifier plays it back, is covered in sample rate and bit depth explained.