Is the filter algorithm more important than the tap count, and would more than a million taps keep improving the sound?

Short version

Watts holds two positions together. First, the algorithm (for him, the WTA windowing function that calculates the coefficients) matters more than the tap count, and a long filter built with a conventional algorithm gains little Head-Fi, Dec 2017Head-Fi, Aug 2022Part-Time Audiophile 30:14. Second, with the algorithm held the same, more taps have sounded better in his tests Head-Fi, Oct 2017Head-Fi, Apr 2024, and since hearing a million taps (first in the Blu Mk2's M scaler, then the Hugo M Scaler) he has said 1M is not the end Head-Fi, Jan 2017Head-Fi, Jan 2018Head-Fi, Nov 2020article, Oct 2026.

Is the algorithm more important than the tap count?

2014 to 2016: Hugo and DAVE

In April 2014, discussing Hugo, he said the WTA algorithm gave roughly the sound of a conventional half-band filter with ten times the tap length, while adding that increasing tap size improves performance "even with using the WTA" Head-Fi, Apr 2014. In July 2014 he described a two-part approach: keep increasing tap length until he could hear no improvement, "and secondly that the FIR interpolation algorithm itself has to be optimized to improve timing and maximize sound quality with a finite number of taps" Head-Fi, Jul 2014.

In June 2015 (DAVE thread) he recounted his late-1990s FPGA work: going from 256 to 2,048 taps with half-band Kaiser filters "gave an enormous sound quality improvement", and then with his first WTA algorithm "a 256 tap WTA sounded much better than 2048 tap half band Kaiser, even though the data is being modified" Head-Fi, Jun 2015. In a 2016 Hifistatement interview he put the WTA's gain at "a tenfold improvement in sound" Hifistatement 1:39, and in talks published in 2018 and 2019 he repeated the 256-versus-2,048 comparison as at least a tenfold improvement MajorHiFi 10:05Passion for Sound 25:10.

2017 to 2019: Blu Mk2 and Hugo M Scaler

In January 2017 he said that when DAVE went from 26,000 to 164,000 taps "the improvement was there but I was disappointed", and that reworking the WTA algorithm and moving from 8FS to 16FS were "both of which was much more important in SQ terms than the tap length" Head-Fi, Jan 2017. Days later he listed tap length alongside "the oversampling rate, the algorithm used, the coefficient quantization, and other factors" as all very important Head-Fi, Jan 2017.

In October 2017 he wrote that the M Scaler "isn't telling us how good a million real 16FS FIR taps sound, but how good the algorithm is", and that "The algorithm becomes even more important as the tap length increases" Head-Fi, Oct 2017. In December 2017 he said "there is not much point in dramatically increasing the tap length if a conventional algorithm is used, as you become algorithm limited (more technically windowing function limited), rather than tap length limited" Head-Fi, Dec 2017, and that "Most algorithms would have little SQ benefit in increasing the tap length, as they are very different to ideal sinc function values" Head-Fi, Dec 2017.

On the Hugo M Scaler thread in July and August 2018 he wrote that "the WTA algorithm is actually much more important than tap length" Head-Fi, Jul 2018; a WTA algorithm is "more important than anything else" when he answered a point about PC upsampling Head-Fi, Aug 2018; and, of a filter a poster cited as 10,000,000 taps whose transition band he judged poor, "It reinforces what I have said since the late 1990s - the algorithm employed is much more important than the tap length" Head-Fi, Aug 2018. In the same post he kept the qualifier that "for a given algorithm (assuming it is scaled accordingly) increasing tap length will improve the sound quality" Head-Fi, Aug 2018.

At CanJam London in July 2018 he said the algorithm itself had to change as taps grew: "as my tap length has increased, I've had to change the algorithm" c w 13:58. In March 2019: "it's pointless to increase the tap length using existing algorithms" Head-Fi, Mar 2019.

2020: the million-tap comparison

In November 2020 he described building 1M-tap filters on his M Scaler test rig with a rectangular window (truncated pure sinc coefficients) and with a Kaiser window, to compare against the 1M-tap WTA by measurement and listening Head-Fi, Nov 2020, "to see if at 1M taps if it still held true that algorithm is the most important thing" Head-Fi, Nov 2020. He reported the rectangular-sinc filter's transition band at about 6 kHz against the WTA's 12 Hz and called it "not fit for purpose"; the 1M Kaiser measured better than the WTA (a 3 Hz transition band) but in listening was worse than the 1M WTA and also lost to DAVE's 164k-tap WTA, and he wrote that "the 1M Kaiser sounds similar to what I expect a 20k tap WTA filter would sound like" Head-Fi, Nov 2020. His conclusion: "I ended up being more convinced that algorithm is vastly more important than tap length" Head-Fi, Nov 2020. On November 27 he put the WTA algorithm at "50 to 500 times more effective in sound quality than conventional filters", up from his earlier estimate of more than ten times Head-Fi, Nov 2020.

In the same November 11 post he said tuning the WTA's variables mattered because "a subtle change in values can give the subjective equivalence of doubling the tap length" Head-Fi, Nov 2020.

2021 to 2026: Mojo 2 and Quartet

In 2021 he applied the point to software upsamplers: "Millions, or billions of taps using the wrong algorithm simply gives you the wrong result" Head-Fi, Dec 2021, and "the key here is the number of taps that are Whittaker-Shannon - NOT the number of taps" Head-Fi, Jun 2021. In January 2022, on Mojo 2, he wrote that "taps are only part of the overall sound quality" Head-Fi, Jan 2022. In February 2022: "the algorithm is way more important than the number of taps" Head-Fi, Feb 2022; in July 2022, of apodizing filters, "It doesn't matter if it's 10 taps or 10 billion taps - apodizing is the wrong approach" Head-Fi, Jul 2022; in August 2022: "so the algorithm is more important than the tap length" Head-Fi, Aug 2022.

For Mojo 2, in September 2023 he said an M Scaler would still improve it "as the longer tap lengths on the M scaler further improves transient timing reconstruction accuracy. But the algorithm is very much more important than the number of taps employed" Head-Fi, Sep 2023.

For Quartet (4 million taps), when the pre-production prototype was shown at Munich in May 2024 he said "That is much more important than the number of taps", referring to the algorithm AVstore 4:40. In interviews and talks published in 2026: "the algorithm is fundamentally more important than the number of taps" Audio Force 8:16; "You could have a billion taps, wrong algorithm and a thousand taps would sound better" Part-Time Audiophile 30:14; and, of Quartet, "So the 4 million taps isn't the story. The story is the new algorithm, the new filter structure" article, Oct 2026.

The qualifier he attaches

He has paired the comparative claim with a same-algorithm rule. October 2017: "for a given algorithm, (so far at least) increasing the tap length has always improved the sound quality", and "For the WTA, doubling the tap length, doubles the accuracy against the ideal" Head-Fi, Oct 2017. A 2017 Hifistatement clip: "with a given algorithm, the more taps that you have, the better the sound quality will be" Hifistatement 1:43. April 2024: "using the same WTA algorithm, and careful listening tests, the doubling of tap length has always led to a corresponding improvement in sound quality" Head-Fi, Apr 2024. In a podcast published October 2026: "with a given algorithm, if you double up the number of taps and make the same adjustments within the algorithm, you will get better sound quality" Part-Time Audiophile 30:58.

In an October 2017 talk he recalled that in his FPGA work from about 1998 a standard Kaiser filter sounded better "every time you double the tap length", and that being unable to get beyond about a thousand taps at that stage is what led him to look at changing the algorithm miceblue425 20:34miceblue425 21:21.

Would more than a million taps keep improving the sound?

2014 to 2016: before a million taps existed

With Hugo in April 2014 he wrote that "there has to be a limit" to when tap size stops improving the sound, guessed 1M taps for an 8-times oversampling filter because the sinc coefficients would then fall below the 16-bit level, and added "But this is a guess. It could be 100K, it might be 10M. Nobody knows." Head-Fi, Apr 2014. Two days later: "At some point though it will cease to make a difference, but I don't know where this point is" Head-Fi, Apr 2014.

In March 2015 he reported still hearing gains from 100,000 to 164,000 taps and wrote "we do not need infinite tap length filters, but we certainly need much more taps than even 164,000 before one can no longer hear the difference" Head-Fi, Mar 2015. In September 2015 he added "albeit I can now start to hear the law of diminishing returns" Head-Fi, Sep 2015. In March 2016 he was less certain: "I suspect that longer tap lengths will give better sound. But by how much? Frankly I do not know, and its possible its not much." Head-Fi, Mar 2016. In October 2016: "Even with Dave at 164,000 taps, we have not reached the subjective limit yet" Head-Fi, Oct 2016.

2017 to 2019: after hearing the M Scaler

When the M Scaler technology appeared in the Blu Mk2 in January 2017, he said he had expected half a million taps to be only a worthwhile improvement but found it transformational, and wrote "Now we know that 1M taps is much better than 0.5M taps - so how far away can we go, how deep is the barrel?" Head-Fi, Jan 2017. In the same post, of 256 million taps: "I am sure it won't be necessary, but I have been wrong before about expectations" Head-Fi, Jan 2017. He also wrote that "hopefully the 1M taps is very close to the original, as I really really do not want to spend years coding for ultra long tap lengths" Head-Fi, Jan 2017, and days later that a poster's camera-focus analogy "also implies that 16 bit at 768k is close to ideal - and that is where Davina will tell me for sure" Head-Fi, Jan 2017. He expected his Davina ADC project to settle how close 1M taps was to ideal Head-Fi, Jan 2017Head-Fi, Jan 2017.

He repeatedly said the process was not finished: "even though I am at 1 million taps with the M scaler I doubt that this is the end of the process" (April 2017) Head-Fi, Apr 2017; "I know that going from 0.5M taps to 1M taps is very audible, so this quest is not finished" (January 2018) Head-Fi, Jan 2018; and at CanJam London in July 2018, "It's not job done", with his feeling that "we need to go further" c w 31:47. In July 2018 he also wrote "for sure there must be a point of diminishing returns" Head-Fi, Jul 2018.

In a January 2018 reply (the archived post does not show the question), he wrote that the WTA and M scaler had been so finely tuned "that I am not expecting any further significant improvements", and framed the open question as whether 20-bit accuracy (16,000,000 taps) or 24-bit accuracy (256,000,000 taps) was needed, calling 16M "extremely expensive" and 256M "currently impractical" Head-Fi, Jan 2018. In an October 2017 talk he said 24-bit sinc reconstruction would need 256 million taps, "So that's never going to happen", followed by "I doubt very much that's going to be audible"; the transcript does not make clear what that last remark refers to miceblue425 40:07.

In September 2018 he gave as his "current assumption" that "any transient timing error that is constantly varying is likely to be audible, no matter how small that error is" Head-Fi, Sep 2018.

2020 to 2022

"So with 1M taps are we getting close to subjective convergence? Absolutely not." he wrote in November 2020, adding "I am certain that more musical benefits are to be had by further increasing the tap length" Head-Fi, Nov 2020. Three days later he said the goal is the point where doubling brings no improvement, "but we are not there yet" Head-Fi, Nov 2020. In January 2021: "going from 164k WTA taps to 1M WTA taps was a larger change subjectively than 26k WTA to 164k WTA" Head-Fi, Jan 2021. In July 2021 he wrote that "My recent research is indicating that we need much better accuracy than 16.7 bits - the ear/brain is extremely sensitive to vanishingly small timing errors" Head-Fi, Jul 2021.

He added one limit in January 2022: "as you go higher in sample rate the benefits from increased tap length diminish markedly", adding that "768 kHz does seem to be the sweet spot for recordings", so future M Scalers would stay at 705/768 kHz Head-Fi, Jan 2022. In an interview about the Hugo M Scaler published in August 2022 (recording date not stated) he said "now I know that 16 bit accuracy is not good enough" and called it "a never ending situation in terms of how accurate we need the reconstruction of transients to be" AVstore 6:55. In the same interview, about the scaler then in development for DAVE: "So I already know that more taps gives you better performance. I must stress it's not just about the taps, it's about the algorithm." AVstore 7:51. In January 2023: "increase the WTA tap length further and perception of depth still increases" Head-Fi, Jan 2023.

2023 to 2026: the count he now cites, and Quartet

He has described research, begun in 2018 when the M Scaler was finished, into how many taps are needed article, Oct 2026article, Oct 2026. At CanJam Singapore in March 2023 he said the answer was "billions upon billions upon billions" of taps, "ten to the power of seventeen", equal to a four-billion-year filter, and that he had found "a way of improving that" Eric's Hi-Fi 4:45Eric's Hi-Fi 5:21.

Quartet, first shown as a pre-production prototype in May 2024 and launched in 2026 (in a talk published March 2026, recording date not stated, he spoke of being there "so that we can finally launch Quartet"), has 4 million taps, which he described as four times the Hugo M Scaler's tap length AVstore 4:31Hifi Pig TV 1:28Audio Force 4:40. In July 2024 he repeated the 16-bit point: "The 1M tap M scaler gave 16 bit reconstruction accuracy, which is correct; but that doesn't tell you whether 16 bit accuracy is good enough. I now know that any transient timing error, no matter how small, is perceptually significant" Head-Fi, Jul 2024, and in a Korean interview he said of Quartet's development, "16 bit is not good enough" HIFICLUB 20:23.

In 2026 he gave a larger figure. In the October 2026 Soundnews interview the brute-force count is "four times 10 to the 28", with a filter delay of "900 trillion years", and "this isn't the best. This is getting to a good reproduction ability" article, Oct 2026article, Oct 2026. He said two contrary effects degrade transient timing in any single-stage filter, so improving one worsens the other article, Oct 2026, and that separating them led to Quartet's Blackbird WTA structure, which brings the requirement from 900 trillion years to "tens of seconds" article, Oct 2026article, Oct 2026. talks published in August and October 2026 give the same account, with the count heard as "28 zeros after the four" in one and "four with twenty-six zeros" in the other The Source AV 7:27Distorsiuni Interioare 19:19.

Asked whether Quartet's 4 million taps should increase, he said yes: "Yes, still. Because I think the limit is now tens of seconds, and that will be a project in the future, to do an Ultima scaler" article, Oct 2026article, Oct 2026. He said 768 kHz recordings from his pulse array ADC, decimated and re-interpolated, would show the losses, and "The objective is to get the difference subjectively so small that we don't need to worry about it" article, Oct 2026. In the same interview: "The algorithm itself is much, much more important than the number of taps, as the Blackbird illustrates and as the WTA illustrates" article, Oct 2026.

What he has not said

His only tap counts for where improvement would stop are brute-force estimates for a conventional single-stage approach: the 2023 talk's ten to the power of seventeen, "before we no longer need to worry about it", and the 2026 figure of four times 10 to the 28, of which he said "this isn't the best" Eric's Hi-Fi 4:45article, Oct 2026article, Oct 2026. He has given no tap count at which the Blackbird approach would stop improving: for it he gives a filter duration, "tens of seconds", rather than a tap figure article, Oct 2026. This archive also holds no reported result from the decimate-and-reinterpolate comparison he has described planning since 2016 Head-Fi, Mar 2016Head-Fi, Jan 2017article, Oct 2026.

Quotes

“But of course, for a given algorithm, (so far at least) increasing the tap length has always improved the sound quality”2017-10-28 Head-Fi, Oct 2017

“So with 1M taps are we getting close to subjective convergence? Absolutely not.”2020-11-11 Head-Fi, Nov 2020

“You could have a billion taps, wrong algorithm and a thousand taps would sound better.”2026-10-14 Part-Time Audiophile 30:14

“Yes, still. Because I think the limit is now tens of seconds, and that will be a project in the future, to do an Ultima scaler.”2026-10-17 article, Oct 2026

Dated positions

Caveats and open points

• The brute-force tap counts differ by source: a 2023 talk gives ten to the power of seventeen; the edited Soundnews interview (2026) gives four times 10 to the 28; 2026 talks were heard as 28 zeros after the four and as four with twenty-six zeros. Check the recordings before relying on any of these figures.
• In an October 2017 talk transcript, his remark that he doubted something would be audible follows the 256-million-tap figure, but the transcript does not make clear what it refers to.
• The tenfold and 50 to 500 times figures are his own listening judgments, and the 2020 rectangular-sinc and Kaiser comparison is his own description of his own tests.
• Tens of seconds is a filter duration for the Blackbird approach, not a tap count; he has given no tap figure for that target.
• Statements are scoped to the products named (Hugo, DAVE, Hugo M Scaler, Mojo 2, Quartet); his 2022 remark that tap-length benefits diminish at higher sample rates was about recordings and future M Scalers at 705.6/768 kHz.
• The 2017 Hifistatement German article text (publisher's report) was not used as his words.
• His January 2017 hopes that 1M taps might be very close to the original, or that 16 bit at 768k might be close to ideal, were left open pending Davina; from 2021 he said 16-bit accuracy is not good enough. These are dated statements, not a declared change of view.

Topics: WTA filter algorithm and coefficient design, Tap length and diminishing returns