Should audio processing dither or noise-shape, and is 64-bit floating point transparent?

Scope. Both topics come up when a digital processing stage has more bits than the next stage accepts and must truncate them, which he calls re-quantization or truncation Head-Fi, Jul 2017. What follows reports his listening results, simulations and design choices as he describes them; none of it is independent measurement.

Is 64-bit floating point transparent?

In the dated items from November 2019 to October 2026 that address it, he says it is not Head-Fi, Nov 2019Head-Fi, Oct 2021Head-Fi, Dec 2021Distorsiuni Interioare 36:27. This archive holds no statement of his on floating point before November 2019 Head-Fi, Nov 2019. The main reason he gives is noise floor modulation, which in 2019 and 2021 he ties to the exponent changing the resolution as the signal level changes Head-Fi, Nov 2019Head-Fi, Jun 2021; for EQ he also cites small-signal amplitude and phase errors Head-Fi, Feb 2022.

In November 2019 (Hugo M Scaler thread) he wrote that "floating point is not a good idea as it creates noise floor modulation" Head-Fi, Nov 2019. In July 2020 he proposed a test for a PC equalizer running at 64-bit floating point: set every parametric slider to -1 dB and raise the DAC volume by 1 dB. He wrote that "A perfect EQ should sound identical with it on or off" but that when he tried it it "clearly does not sound the same", describing his M scaled Dave as sounding brighter with treble glare and flatter depth Head-Fi, Jul 2020. In June 2021 he called the limits of 64-bit floating point "a big problem" and said the noise floor modulation it produces is audible, giving "a bright, slightly hard sound quality" Head-Fi, Jun 2021; later that month he wrote that all of his audio digital modules are better than 64b floating point Head-Fi, Jun 2021.

In October 2021 he wrote "64 bit floating point is not transparent" Head-Fi, Oct 2021, and in December 2021 "64 bit floating point is not accurate enough for transparency - it creates noise floor modulation innately" Head-Fi, Dec 2021. On 31 January 2022, in a technical presentation on Mojo 2, he wrote that "Eliminating noise floor modulation in digital modules is possible with fixed point architecture and using noise shaping for truncation" Head-Fi, Jan 2022. In a podcast the same day he said the modulation "is actually very low, but it's very, very audible" Darko.Audio 20:36. In February 2022 he wrote "64 bit DSP is, in my opinion, inadequate to transparently do EQ", and said a transparent EQ needs a lot more bits than 64, fixed point processing and noise shaping on all internal nodes at 705/768 kHz or higher, which he says no current EQ does except Mojo 2's (104 bits) Head-Fi, Feb 2022. At CanJam NYC 2022 (late February; video published March 2022) he put the modulation at about minus 300 dB but very audible, and said a fully transparent EQ "cannot be floating point" MFG Audio Reviews 34:48MFG Audio Reviews 35:31.

One qualification he gives concerns precision. In July 2023 he wrote that going from 64-bit to 128-bit floating point "will get you closer, and 256b FP almost converges to fixed point noise shaped operation", while saying 64-bit floating point has "very serious subjective consequences" in his opinion Head-Fi, Jul 2023. In November 2023 he added that the errors are "very significant subjectively" to him, and that noise shaping lets a filter keep working on signals below the system's bit depth Head-Fi, Nov 2023.

His most recent statements in this archive are talks. In an interview recorded at a Munich show, most likely May 2024 (date unconfirmed; published February 2025), he said floating point "creates noise floor modulation, and the noise floor modulation is very, very audible", and that 64-bit floating point on a PC is not good enough for accuracy, particularly in IIR work audio-2G 15:44audio-2G 17:15audio-2G 17:33. In September 2024, describing new tools for analyzing his filter coefficients, he said they were no longer limited to Windows 64-bit floating point, which he said "limits everything to 300 dB" The Source AV 8:10. In a podcast recorded in October 2025 (published March 2026) he said "it doesn't matter how small the noise floor modulation is, it is perceivable", preferring the word perceivable to audible ecoustics 48:39. In a podcast published in October 2026, recorded at the Quartet launch event in Romania on a date the archive does not record, he said floating point creates noise floor modulation that is quite small but audible, and that errors compound when processing goes into an FFT and back through an IFFT Distorsiuni Interioare 36:27.

Dither or noise shaping?

2014 to 2015 (Hugo). In September 2014, about JRiver's volume control, he wrote that if it "is properly dithered when truncating, then you do not need to worry about distortion or loss in transparency" Head-Fi, Sep 2014. In January 2015 he said Hugo's 51-bit internal volume result must be truncated, that "the truncation is done via noise shaping and dithering", and that this adds "just a fixed noise at -180dB", which he said was verified with Verilog simulation Head-Fi, Jan 2015.

2016 (Dave, Blu and the Davina project). In March 2016 he wrote that, in principle, properly dithered 16-bit audio can perfectly resolve an infinitely small signal given an infinite period for the FFT or correlation, and that for Dave's 16FS signals truncation is easy with a 350 dB noise shaper, but "this is not an option at 44.1" Head-Fi, Mar 2016. A few days later he wrote that he normally uses noise shapers to guarantee 64-bit audio performance, which works at 768 kHz but "won't work effectively at 44.1 kHz", and asked whether triangular dither could do the job: "In principle it will", with the Davina project to find out Head-Fi, Mar 2016. In May 2016, replying to a report of dither improving sound, he said that adding random noise had helped his older multi-chip designs, but that after pulse array "employing extra dither no longer made an improvement in sound quality", and that on Chord's Blu CD transport it remains an option meant for "poorer DAC's" Head-Fi, May 2016. He added that this was not truncation dither, since the data being randomized was zeroes Head-Fi, May 2016. In October 2016, testing the Blu Mk 2 prototype at 768 kHz, 24-bit, he wrote "The triangular had the worst sound quality - poorer depth, and detail resolution. Gaussian was better, but the noise shaper was easily the best sounding" Head-Fi, Oct 2016.

2017 to 2019 (Hugo 2, Blu 2, TT2, Hugo M Scaler). In April 2017, about EQ done on a PC, he wrote that his tests showed dither "is not as good in preserving depth as doing it at 768k with noise shaping", adding "We are not talking about vast changes though" Head-Fi, Apr 2017. A day later, on the 44.1 kHz case, he wrote "So far its looking like Gaussian dither is the best sounding option" Head-Fi, Apr 2017. In July 2017 he reported his pseudo Gaussian dither as "about half way" to the noise-shaped sound and said he hoped to improve it Head-Fi, Jul 2017. In February 2018, recalling Blu 2's output truncators, he wrote that triangular (TPDF) dither had the worst sound with "a reduction in depth", Gaussian was about half way to the noise shaper, and there were "(small but audible) transparency losses involved in TPDF" Head-Fi, Feb 2018. Later that month, on going from 24 to 16 bits, he wrote that "the only way that I have managed to do this completely transparently (so far) is with extremely aggressive noise shaping at 705/768 kHz", and listed as open questions whether bandwidth limiting and a closer-to-ideal Gaussian dither would matter, which Davina would answer Head-Fi, Feb 2018. In October 2018, about TT2's volume control, he wrote that his advanced noise shaper "sounds much more transparent than just adding dither" Head-Fi, Oct 2018. In December 2018, in the Hugo M Scaler thread after discussing its 4x (192 kHz) filter, he said the dither employed is Gaussian at 24 bits "as this sounds much better than triangular or rectangular dither" Head-Fi, Dec 2018; in March 2019 he said the M scaler adds no dither at all when using the full 1M taps Head-Fi, Mar 2019. In March 2019 he wrote "The only way that is audibly transparent is to use very aggressive noise shaping; second best route is gaussian distribution dither" Head-Fi, Mar 2019, and that changing volume and "re-quantising to 24 bits with triangular dither is certainly not audibly transparent" Head-Fi, Mar 2019.

In December 2019, asked whether the M scaler dithers down to 24 bits, he answered "it's not dithered down to 24 bits but very aggressively 11th order noise shaped". He said rectangular dither sounded worst, triangular was "still much inferior" to 768 kHz noise shaping, and pseudo Gaussian was much closer but "no where near as good as the noise shaped performance". His advice was to use Gaussian dither if a source offers it, otherwise triangular Head-Fi, Dec 2019.

2020 to 2021. In February 2020 he wrote "Dithering to restore the bit depth is not transparent" and that the only way he has done it is at 705/768 kHz with aggressive 11th order noise shaping Head-Fi, Feb 2020. In March 2020, about Qutest (which has no volume control of its own), he wrote that "just truncating by discarding the unused bits creates distortion so don't do that - but dither isn't perfect", that "The best from a transparency POV is Gaussian dither, which is better than the usual triangular dither", and that "The only way of completely maintaining full transparency when truncating is to use 11th order noise shaping, and this only at 705/768 kHz"; he still advised software volume in Roon over a preamp or passive control Head-Fi, Mar 2020. In July 2021, about his volume control (Dave thread), he wrote that its output "is simply not truncated - nor is it dithered - but aggressively noise shaped", and that digital volume controls are normally dithered when truncating, "which is subjectively not lossless" Head-Fi, Jul 2021. In August 2021 he qualified this: his subjective evidence so far indicates his truncation noise shapers are lossless from a sound quality point of view, but calling them absolutely lossless would be technically inaccurate, and he was researching how far noise-shaped truncation can go before it stops being subjectively transparent Head-Fi, Aug 2021. In October 2021 he wrote "Absolutely you must always dither, but the best approach is noise shaping", and put pseudo Gaussian at "only half as effective subjectively as the truncation noise shaping"; he added that so far the evidence suggested transients are not degraded by dither or noise shaping at the levels he uses, which he was still examining Head-Fi, Oct 2021.

2022 to 2024. In March 2022, asked about a source's EQ dithering, he said to try 24-bit and 32-bit outputs and use whichever gives the best impression of depth Head-Fi, Mar 2022. In July 2022 he explained that the Hugo M Scaler uses Gaussian dither "set to a high level" for 2x and 4x upsampling because it sounds closer to the noise-shaped 16x mode, and that at 16x the noise and distortion equal the input's Head-Fi, Jul 2022. He said the single BNC output adds some noise because it uses Gaussian dither, which he called inaudible Head-Fi, Jul 2022. On 7 July 2022 he cited the Lipshitz, Wannamaker and Vanderkooy dither paper to argue that rectangular dither is rejected for noise modulation, that triangular dither does not cover higher moments of the error, and that in his listening the order was rectangular worst, then triangular, then pseudo Gaussian, with a "noticeable degradation" still remaining against the noise-shaped reference Head-Fi, Jul 2022. In July 2022 he said that 24-bit re-quantization at 48 kHz "completely transparently is a problem waiting to be solved" Head-Fi, Jul 2022. In January 2023 he wrote that a 192 kHz output "for measurements with the AP" uses pseudo Gaussian dither with higher noise than triangular but a sound much closer to the 768 kHz noise-shaped output; the post does not name the product, and his next post in the thread discusses the Hugo M Scaler's 24-bit outputs Head-Fi, Jan 2023Head-Fi, Jan 2023. In June 2024, his latest dated comparison of dither types in this archive, he wrote that Gaussian is technically better than triangular (his pseudo Gaussian uses around fifty random sources) and repeated the listening order: rectangular worst, then triangular, then pseudo Gaussian, with noise shaping best, the difference being depth perception Head-Fi, Jun 2024.

2025 to 2026. In a podcast recorded in October 2025 (published March 2026) he said information below the 16-bit level gets encoded "by the use of dither or by using noise shaping", and that if a signal is properly dithered very small signals come through ecoustics 7:02ecoustics 7:56. In the same podcast, on purely digital truncation noise shapers (digital in, digital out), he said listening tests gave the same result as for his other noise shapers, needing better than 350 dB performance; the transcript wording is garbled at the end of the sentence, so check the recording ecoustics 39:56ecoustics 40:09.

How firmly he holds this

He describes the findings as listening results. In June 2017, discussing listening tests on noise shaper performance levels (not dither types), he acknowledged that they were not double blind and that confirmation bias could not be ruled out, while writing "Because of these multiple tests, I am convinced that these results are real" and planning a blind test with Davina recordings Head-Fi, Jun 2017. In 2021 he wrote of his truncation noise shapers that "All the subjective evidence I have to date indicates" they are lossless, and described a research program to test the claim directly Head-Fi, Aug 2021.

What he has not said. He does not say dither can be skipped when truncating; his 2020 and 2021 posts say truncating without it causes distortion Head-Fi, Mar 2020Head-Fi, Jul 2021. He does not give measured results for Gaussian versus triangular dither in these items, only listening rankings and descriptions such as "about half way" and "small but audible" Head-Fi, Feb 2018. His full-transparency noise-shaping claims rest on the 705/768 kHz case: in 2019 he said he still had more work to do on 44.1 kHz truncation to 24 bits, in July 2022 he called transparent 24-bit re-quantization at 48 kHz unsolved, and in November 2023 he said fully transparent EQ at 44.1 kHz "could be a major headache" Head-Fi, Dec 2019Head-Fi, Jul 2022Head-Fi, Nov 2023.

Quotes

“Absolutely you must always dither, but the best approach is noise shaping - but you need to be running at 705/768 for this to be 100% effective.”2021-10-05 Head-Fi, Oct 2021

“The best from a transparency POV is Gaussian dither, which is better than the usual triangular dither”2020-03-09 Head-Fi, Mar 2020

“Dithering to restore the bit depth is not transparent; the only way I have done it is with 705/768 kHz and to use aggressive 11th order noise shaping”2020-02-22 Head-Fi, Feb 2020

“64 bit floating point is not accurate enough for transparency - it creates noise floor modulation innately.”2021-12-07 Head-Fi, Dec 2021

Dated positions

Caveats and open points

• His dither comparisons are reported listening results. In Head-Fi, Jun 2017 (2017) he said his listening tests on noise shaper performance levels were not double blind and confirmation bias could not be ruled out; that post concerns noise shapers, not dither types. This archive contains no blind test of dither types.
• Statements are scoped to the product or stage he names: the M scaler (16x noise shaped; 2x, 4x and single BNC Gaussian dither), Dave's and TT2's volume controls, Qutest, and the Blu, Blu Mk 2 and Davina projects. Head-Fi, Jan 2023's 192 kHz measurement output is not named in the post. A statement about one product does not carry to the others unless he says so.
• His 2014 post about JRiver's volume control (Head-Fi, Sep 2014) treats properly dithered truncation as nothing to worry about, while his 2019 to 2021 posts say dithering is not audibly transparent; this archive does not show him reconciling the two, so it is reported as a difference between posts, not as a stated change of view.
• His 'pseudo Gaussian' dither is his own implementation; he gave 14 random sources in 2017, planned 56, and described about fifty in 2024 (Head-Fi, Jul 2017, Head-Fi, Jun 2024). Posts do not say which count a given product uses.
• Head-Fi, Dec 2019 contains a gap in the forum text (a […] where the scraper dropped words) in his advice on a filter truncation setting; the dither-off remark is not quoted for that reason.
• The 2016 posts Head-Fi, May 2016 and Head-Fi, May 2016 concern randomizing data at the DAC input, which he says is not truncation dither (Head-Fi, May 2016); they are reported here for completeness.
• Dating of the talks: the audio-2G interview was recorded at a Munich show, most likely May 2024 (date unconfirmed), and published February 2025; the ecoustics podcast was recorded 2025-10-28 and published March 2026; the Romanian podcast was published October 2026 and recorded at the Quartet launch event, on a date the archive does not record; the CanJam NYC 2022 video was published March 2022 but, by the source record, the talk was in late February (show dates unverified).

Topics: Re-quantization, truncation, dither and word length, EQ, tone controls and room correction