What is noise floor modulation, how does Rob measure it, and why does he say it is audible below what can be measured?

What he means by noise floor modulation

Watts defines noise floor modulation as a change in a device's noise level with signal level. In June 2015 he gave the example of a DAC whose noise is -120 dB with a -60 dB sine but -115 dB with a 0 dB sine, so that "the noise floor (seen by doing an FFT measurement) would increase by 5dB" Head-Fi, Jun 2015. He has given the same definition since, most recently in an interview published in October 2026 (recording date not stated; its transcript places it before May 2026), describing the ideal as a noise floor that stays the same as the signal rises and falls Dusty Audio 31:39HiFi Buys 52:42. He separates it from fixed noise, which he calls unimportant background hiss, like an air conditioning system in a concert hall Pursuit Perfect System 9:36MFG Audio Reviews 5:54. In March 2018 he said it can appear as "the whole noise floor pumping up and down, skirts on fundamentals, or certain bands, or noise floor mod that is slew rate related" Head-Fi, Mar 2018. In 2015 and 2016 he described low-frequency jitter skirting around a tone as an effect that sounds like noise floor modulation but is distinct from it, saying in September 2016 that "normal noise floor modulation is the whole noise floor 20 Hz to 20 kHz pumping up and down"; his March 2018 list above counts skirts on fundamentals among the ways it can be seen Head-Fi, Oct 2015Head-Fi, Sep 2016Head-Fi, Mar 2018.

What he says causes it

The cause he names most often is random RF noise intermodulating with non-linear analog stages, which puts random noise into the audio band that rises with the signal; he traces this to mains-cable listening tests in the 1980s Head-Fi, Jun 2014Head-Fi, Aug 2016Head-Fi, Sep 2022ecoustics 58:44. He distinguishes random RF noise, which he says creates noise floor modulation, from fixed-frequency switching noise, which he says does not Head-Fi, Sep 2020Head-Fi, Oct 2021Eric's Hi-Fi 16:50. Inside DACs he names the reference voltage in differential chip DACs, whose noise cancels for small signals but not large ones Head-Fi, Aug 2014Head-Fi, Aug 2016; signal-dependent switching and glitch energy in R2R, DSD and current-source DACs Pursuit Perfect System 16:35HiFi Buys 1:11:40Head-Fi, Oct 2015; DSD modulator instability Head-Fi, Nov 2018; floating-point DSP Head-Fi, Jan 2022Darko.Audio 20:36; and out-of-band ADC noise-shaper noise in high-sample-rate recordings Head-Fi, Jan 2018. He credits pulse array's constant switching activity for his DACs avoiding measurable noise floor modulation Head-Fi, Mar 2016HiFi Buys 1:12:47.

How he measures it

Before late 2014 he fed an Audio Precision analyzer through a fixed-frequency passive notch filter, but he said that analyzer "was not good at resolving small noise floors" Head-Fi, Jun 2015. He bought an APx555 after its launch around October 2014 and describes it as using one ADC for the fundamental and another fed through an automatic analog notch, with the two results stitched together digitally, which he says "neatly resolves the noise floor modulation issue from ADC's" Head-Fi, Jun 2015Head-Fi, May 2018Head-Fi, Jan 2022. He sometimes describes this as two ADCs per channel and sometimes as four ADCs or a "quad ADC structure" Head-Fi, Mar 2022Head-Fi, Feb 2018Head-Fi, Nov 2021. He says this matters because "All ADCs have very large levels of noise floor modulation," and in March 2022 he told readers "So ignore all measurements unless this is done" Head-Fi, Mar 2022.

His stated procedure (2021 onward) is to overlay a full-scale or 2.5 V RMS tone against the signal turned off, "where you overlay 0dBFS with a -999dBFS signal, with all settings on the APx555 being identical" Head-Fi, Nov 2021Head-Fi, Nov 2021. He adds that input ranging must be fixed (for example at 2.5 V) and "not auto-ranging, otherwise you will see false noise floor modulation", that the high-resolution sine analyzer must be fixed to the fundamental frequency, and that "You can't use the AP defaults to do this test" Head-Fi, May 2022Head-Fi, Sep 2022Passion for Sound 10:05HiFi Buys 53:19. He uses FFTs of about one million points, which lowers the displayed floor to near -180 dB Head-Fi, Mar 2016Head-Fi, May 2018, and in 2020 said the noise floor modulation tests take about an hour Head-Fi, May 2020. In March 2016 he also reported a band-limited, A-weighted THD and noise reading for DAVE: 2.6 uV with no signal, with a -60 dB signal and with a 2.5 V output, a result he called "completely unvarying" Head-Fi, Mar 2016.

He names limits of the method. Because the APx555 uses one notch, "you can only use a single tone sine wave input" Head-Fi, Jan 2022. In the interview published October 2026 (recording date not stated), describing the reference voltage in chip DACs and switching in R2R and DSD DACs, he said "Now when you do static measurements, you're not going to see the dynamic modulation effect," which he said appears only when the sine wave is switched on and off HiFi Buys 1:10:51HiFi Buys 1:11:40.

For digital modules he measures by Verilog simulation and FFT, saying in 2018 that "a Verilog simulation is not a model or an estimation; it is the actual performance of the digital module," and that it shows aberrations "to a performance level one could never measure on a real instrument" Head-Fi, Apr 2018Head-Fi, Apr 2022. This is how he arrives at noise floor modulation figures far below -180 dB Head-Fi, Jun 2015Head-Fi, May 2018.

What he reports for his products

DAVE: from June 2015 on he has published APx555 plots with a 2.5 V RMS 1 kHz trace and a no-signal trace showing the same floor near -180 dB, concluding "there is no measurable noise floor modulation at all" Head-Fi, Jun 2015Hifistatement 22Pursuit Perfect System 18:40. In August 2021, in a post about remeasuring the DAVE unit Stereophile had reviewed, he reposted other DAVE plots and said they show "zero measurable noise floor modulation", and in September 2022 he said it had been "independently confirmed by GoldenOne" Head-Fi, Aug 2021Head-Fi, Sep 2022.

Other products: in 2014 he showed Hugo with a constant -155 dB floor and no noise floor modulation for signals from -20 dB to -140 dB, a small-signal test rather than a full-scale one Head-Fi, Jul 2014. In August 2016 he said "Even Mojo has no measurable noise floor modulation at -170 dB" Head-Fi, Aug 2016. For Mojo 2 he showed a -176 dB floor at 2.5 V and with no output in November 2021, and cited -177 dB at CanJam NYC 2022 Head-Fi, Nov 2021MFG Audio Reviews 45:02. He published Hugo 2 plots in 2017 and 2018 Head-Fi, Mar 2017Head-Fi, Feb 2018 and TT2 plots from 2018 to 2020, including 12 W into 8 ohms Head-Fi, May 2018Head-Fi, Mar 2019Head-Fi, Jul 2020. In a talk published May 2026 (recording date not stated) he said Cortax has no measurable noise floor modulation Audio Force 14:34. He describes "no measurable noise floor modulation" as a sign-off requirement for every design Head-Fi, Jan 2018Head-Fi, Feb 2018Pursuit Perfect System 41:18.

For comparison he has said conventional DACs rise from about -160 dB to about -140 dB, "so that's a 20 dB shift," with class D amplifiers worse RMAF 10:38Pursuit Perfect System 11:05, and in 2016 put typical class D noise floor modulation at "say around -120 dB" Head-Fi, Aug 2016. In the interview published October 2026 (recording date not stated) he gave a smaller figure, that "high performance DACs will have two or three dBs worth of noise floor modulation" HiFi Buys 53:19. In 2017 he said a standard ADC has "about four or five dB's worth" miceblue425 39:02.

His wording has changed. In 2016 and 2017 posts he sometimes wrote of "zero noise floor modulation" for pulse array or his DACs Head-Fi, Mar 2016Head-Fi, Aug 2017, though in July 2016 he already said DAVE has "immeasurably small amounts of RF noise modulation" Head-Fi, Jul 2016. From late 2018 on he stresses the qualifier: "I say measurable because they still have some noise floor modulation" (November 2018 open-day talk), and "I've not succeeded in eliminating noise floor modulation" (July 2024) Pursuit Perfect System 18:40Passion for Sound 7:31MFG Audio Reviews 8:20HIFICLUB 34:29. In June 2021 he named the reference voltage as a "source of (unmeasurable) noise floor modulation", which he said he mitigates in TT2 and DAVE with the lowest-noise references, at levels "below measurement at 150 pV" Head-Fi, Jun 2021.

Why he says it is audible below measurement

His perceptual explanation, consistent from 2014 to talks published in 2026, is that "noise floor modulation is white noise pumping up and down with the signal level," that the brain cannot separate that noise from the instrument it tracks, so warm instruments such as a saxophone sound brighter, and that noise tracking several instruments at once blurs instrument separation and causes listening fatigue Head-Fi, Jun 2014Passion for Sound 8:51MFG Audio Reviews 14:45HiFi Buys 56:59. In a talk published May 2026 (recording date not stated) he said it matters more than harmonic distortion because "the brain is used to dealing with harmonic distortion" and "The problem with noise floor modulation is it's not harmonically related" Audio Force 12:18. He also says the noise change itself would not be heard without the instrument, and in a podcast recorded in October 2025 (published March 2026) said he prefers "perceivable" to "audible": "I'll use the term perceivable rather than audible because you wouldn't hear the error" HiFi Buys 54:03ecoustics 48:39.

On the measurement side, he says real measurement stops near -180 dB on DAVE ("currently we can measure noise floor modulation at -180 dB"), while in 2018 he said that on DAVE "the noise change with signal is unmeasurable" Head-Fi, Aug 2016Head-Fi, Apr 2018Head-Fi, Oct 2020. Below that he relies on digital-domain simulation, which in 2023 he said is the only way to measure such errors: "You can only measure it with digital test gear" Eric's Hi-Fi 31:00.

The evidence he gives is of two kinds. First, simulation plus listening: in 2015 he said he had listened to noise floor modulation at around -200 dB, with figures "derived from simulation," and heard the effect when the mechanism was switched on and off Head-Fi, Jun 2015Head-Fi, Apr 2018MFG Audio Reviews 15:45. Second, audible changes with no measured change: adding USB RF filters from Hugo to Hugo 2, after which, in the Passion for Sound interview (published in parts December 2019 to January 2020; recording date not stated), he said "I know non-measurable noise floor modulation is still significant" Passion for Sound 12:11; DAVE's USB and optical inputs, where in October 2020 he said his own listening tests "revealed a barely audible tiny change" while "USB has identical measured noise floor modulation as optical" Head-Fi, Oct 2020; and his HF filters on 44.1 kHz material, where in 2016 he said "I very strongly suspect it's the noise that is the problem, but I intend to find out for sure" and planned to check for a measured change, and in 2020 ran a single-blind test with his son that he said proved noise floor modulation was the cause Head-Fi, Jan 2016Head-Fi, Apr 2020. In September 2022 he said of analog RF filters, "Most of the time though, it's audible but not measurable," and that listening tests with the Mojo 2 UHD DSP HF filter had "proved (at least to my satisfaction)" that the differences he had heard for decades from analog RF filters were noise floor modulation and not amplitude or phase effects Head-Fi, Sep 2022. He links cable and USB-cable differences to the same unmeasurable effect Eric's Hi-Fi 32:01MFG Audio Reviews 15:45.

The figures he gives have fallen over time, though alongside them he has said since 2015 that there is no lower limit: "There is virtually no limit to how small noise floor modulation needs to be" (November 2015) and "it does not matter how small it is" (March 2018) Head-Fi, Nov 2015Head-Fi, Mar 2018. His figures:

• 2014 to 2020: about -200 dB. He said in June 2014 he had heard noise floor changing at -200 dB Head-Fi, Jun 2014; in December 2015 that "better than 200 dB is a good rule of thumb for noise floor modulation, but certainly not for depth" Head-Fi, Dec 2015; in August 2016 that his "tentative conclusions (or rule of thumb)" put the limit at -200 dB, "and perhaps even lower" Head-Fi, Aug 2016; and in May 2020 that -180 dB to -200 dB was easily audible Head-Fi, May 2020.
• April 2021: "the audible limit is zero," from comparing digital-domain measurements to listening tests Head-Fi, Apr 2021.
• July 2021: interpolator tests suggested "the target for noise floor modulation needs to be around -400dB" Head-Fi, Jul 2021.
• October 2021: "even noise floor modulation as low as -350dB is perceivable under single blind conditions" Head-Fi, Oct 2021.
• 2022 onward: "no effective limit" (September 2022), -350 dB examples in 2023 and in the interview published October 2026 (recording date not stated), and, at CanJam NYC in late February 2022 (video published March 2022), a transcribed remark that he had "heard noise floor at minus 400 dB" Head-Fi, Sep 2022Dusty Audio 34:43HiFi Buys 57:34MFG Audio Reviews 34:48. In May 2024 he said "it doesn't matter how small the noise floor modulation is, even if it's below measurable limits" Reactcore 0. In July 2024 he declined to give his newest figure: "I'm not going to give you the number because the number is much smaller than minus 300 dB" HIFICLUB 48:04. The transcript figures should be checked against the recordings.

On why the brain could be this sensitive, he hedges. In 2024 he said "Honest answer: I don't know," and guessed that the brain uses an autocorrelation function that can resolve very small errors in noise over time MFG Audio Reviews 16:30. He gave the correlation idea as probable in March 2016 ("the brain probably uses correlation routines to separate sounds out") and likely in April 2022, and in a podcast recorded in October 2025 (published March 2026) stated it without a hedge: "this is definitely a case of correlation" Head-Fi, Mar 2016Head-Fi, Apr 2022ecoustics 1:01:20.

He also warns about his own method. Added noise floor modulation sounds brighter and can be mistaken for better transparency, which he says is sometimes almost impossible to tell apart on an AB test Head-Fi, Mar 2016RMAF 2:55. In 2022 he wrote "You have discovered a post where I made a mistake," explaining that a brighter USB input on DAVE in 2016 was later attributed to RF-induced noise floor modulation, and in 2024 described the same error with 2 Qute Head-Fi, Mar 2022Head-Fi, Nov 2024.

What he has not said: in April 2022 he said "I haven't seen anything about noise floor modulation in the psychoacoustic literature" and that he has not published his results (for example in the AES): "It's just my listening tests that suggests that this is a serious problem", and he gives no measured figure for the residual noise floor modulation he says remains in his DACs, only that it is below what the APx555 resolves Head-Fi, Apr 2022Head-Fi, Jun 2021Pursuit Perfect System 18:40.

Quotes

“RF noise, when it is mixed inside the active analogue electronics, creates inter-modulation distortion which results in the noise floor moving up and down with the analogue signal.”2014-06-22 Head-Fi, Jun 2014

“…you overlay 0dBFS with a -999dBFS signal, with all settings on the APx555 being identical. This is the only way to measure noise floor modulation…”2021-11-09 Head-Fi, Nov 2021

“My own tentative conclusions (or rule of thumb) are that one can hear levels of noise floor modulation down to -200dB…”2016-08-05 Head-Fi, Aug 2016

“…even noise floor modulation as low as -350dB is perceivable under single blind conditions.”2021-10-04 Head-Fi, Oct 2021

Dated positions

Caveats and open points

• The RMAF 2017 masterclass transcript says most DACs reach a noise floor of 'about minus 60 dB' at the -60 dB test; this looks like a transcription error and is not relied on here.
• He describes the APx555's ADC arrangement both as two ADCs per channel and as four ADCs; the answer reports both descriptions.
• His sub-measurement thresholds rest on his own listening tests and simulations, which he says are tentative and unpublished.
• The 2014 Hugo result was a small-signal test from -20 dB to -140 dB, not the full-scale versus no-signal overlay he later standardized on.
• The German Hifistatement text is the publisher's translation and is not quoted; the English audio of the same interview is cited instead.
• His figure for high-performance DACs in the interview published October 2026 (two or three dB) is smaller than his 2017 and 2018 figure for conventional DACs (about 20 dB); he does not reconcile the two.
• The draft presented his threshold as moving from -200 dB to zero; he also said 'virtually no limit' in November 2015 and 'it does not matter how small it is' in March 2018, so the no-limit view coexisted with the -200 dB rule of thumb before 2021.

Topics: Noise floor modulation, Measurements, test equipment and simulation, Measurement vs listening