Decimation, aliasing and sample-rate conversion

Reconstruction filters and timing

Answers on this topic

How do microphones, studio conversion and mastering limit what a DAC can reproduce?
Why are half-band decimation filters and studio sample-rate conversion a problem, and how much aliasing rejection is enough?

Everything he said on this topic (115)

One-line summaries in our words, oldest first. Each link opens the original.

2014-04-27Watts explains why native DSD inside a DAC is a bad idea, recounting PDM1024 and pulse array history, and details Qute's DSD filter and Hugo's 140 dB decimation filter. Head-Fi, Apr 2014
2014-06-26Watts advises sending 24/192 natively to Hugo despite its ADC noise-shaper noise, warning that PC sample-rate conversion degrades timing; one PC converting to 48 kHz sounded awful. Head-Fi, Jun 2014
2014-09-30Watts advises using USB over optical rather than letting the PC resample, and says decimating 88.2 to 44.1 is easier than upsampling as decimation has no timing problems. Head-Fi, Sep 2014
2015-02-14Watts says ADC anti-aliasing filters need about 2K taps for -144 dB, ADC chip aliasing near 21 kHz is audible, and his pulse array ADC project has run five years. Head-Fi, Feb 2015
2015-03-04Watts says FIR filters have no simple dB/octave figure, more taps narrow the gap to hi-res, but poor SRC from 96 kHz masters cannot be undone, so recordings vary. Head-Fi, Mar 2015
2015-05-20Watts says DSD is converted to PCM without decimation using improved filters, answering earlier criticism of his decimating approach, which he says had sounded better than previous filters. HIFICLUBE.NET JVH 3:20
2015-05-20Watts says decimation is fine if followed by a WTA filter, but 140 dB decimation accuracy proved insufficient: non-decimating versions sound noticeably better and measurements show distortion at -140 dB. HIFICLUBE.NET JVH 3:52
2015-05-20Watts says next year's pro audio A to D converter needs decimation filters better than 140 dB, while some pro audio filters attenuate only 6 dB. HIFICLUBE.NET JVH 4:50
2016-01-23Watts says the Remedy re-clocker is a chip sample rate converter to 96k, and such SRCs add distortion, noise floor modulation and timing damage, a bad idea with his DACs. Head-Fi, Jan 2016
2016-02-26Watts says Davina, his ADC project prototyped in 2001, tackles noise floor modulation, aliasing (half-band filters give only -6 dB; he hears it at -140 dB) and noise shaper resolution. Head-Fi, Feb 2016
2016-04-01Watts lists what Davina should answer (aliasing limits, lifelike depth with 350 dB noise shapers, transparent WTA round trips, 16-bit dither) and says the power amp will initially matter most. Head-Fi, Apr 2016
2016-04-20Watts says his ADC runs at 104.25 MHz so analog anti-alias filtering is easy, but he does not know how much decimation rejection suffices (-140 dB is not). Head-Fi, Apr 2016
2016-04-24Watts says the best ADC chips manage -122 dB SNR while Davina targets 135 dB, naming depth (350 dB noise shaping), noise floor modulation and audible aliasing as main problems. Head-Fi, Apr 2016
2016-05-17Watts says an ideal sinc reconstruction filter adds no ringing, but theory says nothing on how an ADC should bandlimit; Davina will test aliasing rejection beyond 140 dB. Head-Fi, May 2016
2016-05-17Watts says ringing never explained differences he heard tuning WTA, an ideal filter must ring, 8FS and 16FS filters with identical ringing sound different, and ADC ringing likely matters little. Head-Fi, May 2016
2016-06-06Watts says his ADC plus DAVE at 768 kHz could beat an analog pre-amp, and describes a WTA FIR decimation filter from 104 MHz to 768 kHz for two prototypes. Head-Fi, Jun 2016
2016-09-19Watts outlines Davina test recordings: two prototypes on one mic feed to compare configurations for depth, home and village sounds, then pro engineers, with 768 kHz originals for decimation testing. Head-Fi, Sep 2016
2016-09-19Watts plans Davina test recordings of guitar, fireworks and wood blocks, says current ADCs decimate crudely with moving averages, and plans Earthworks QT50 omni mics with a Schneider disc. Head-Fi, Sep 2016
2016-10-18Watts says he has bottomed out depth, needing -350 dB capable paths with no small-signal linearity issues, but expects more timing discoveries with Davina, especially in ADC decimation. Head-Fi, Oct 2016
2016-12-21Watts says the Remedy's cheap TI SRC chip is not bit perfect, adds distortion, noise and timing errors, and is possibly the fastest way to destroy Mojo or Dave's sound. Head-Fi, Dec 2016
2017-01-29Watts says decimating 192k to 48k is relatively easy, but 192k to 44.1k needs M scaling to 104.25 MHz then decimating, and non-integer SRC is hard to do transparently. Head-Fi, Jan 2017
2017-02-13Watts calls it the 64,000 dollar question, saying that with timing problems removed on the decimation and interpolation side it should sound pretty special. needs context Head-Fi, Feb 2017
2017-02-14Watts says the likely lesson is how far decimation and interpolation are from transparency, and doubts Davina-sourced recordings would change the WTA algorithm, though future gains may follow. needs context Head-Fi, Feb 2017
2017-04-11Watts says Radio 3 passes through three lossy sample rate converters, and recommends setting 44.1 kHz for Radio 3 since most music is 44.1 kHz. Head-Fi, Apr 2017
2017-05-21Watts says down-sampling creates transient timing errors through aliasing, but up-sampling creates many times more, as reconstruction is much harder than decimation. Head-Fi, May 2017
2017-08-11Watts praises un-remastered 1960 to 1967 Decca recordings, says modern ADCs' half-band filters allow aliasing that degrades transient timing, and says Davina will have no aliasing. Head-Fi, Aug 2017
2017-08-12Watts says Davina will measure the audible losses of decimation and re-interpolation, recalls underestimating the M scaler's gain, and expects 44.1/16 Davina recordings to beat conventional ADCs. Head-Fi, Aug 2017
2017-08-13Watts says Davina will test how many taps are enough and whether ADC bandwidth limiting matters, noting sinc reconstruction is required and his IIR decimation filter gives 300 dB rejection. Head-Fi, Aug 2017
2017-08-21Watts says only a sinc reconstructs correctly, short filters sound soft and warm, and he will post blind Davina files comparing FIR and IIR decimation filters with 300 dB attenuation. Head-Fi, Aug 2017
2017-08-21Watts says only a sinc filter perfectly reconstructs sampled data, but bandwidth limiting before sampling can use any filter, provided there is no output at or above FS/2. Head-Fi, Aug 2017
2017-10-11Watts says aliasing from modern ADCs' half-band filters (-6 dB at FS/2) is irreparable, yet every track in his 5 TB library sounds better on BluDave with longer taps. Head-Fi, Oct 2017
2017-10-13Watts says Davina differs from conventional ADCs for three reasons; the first is aliasing, since standard ADCs decimate 1-bit or 5-bit 3 MHz output with crude moving-average sinc filters. miceblue425 36:03
2017-10-13Watts says these decimation filters attenuate only about 20 dB, and the resulting aliasing changes the timing of transients audibly; his second Davina point concerns depth. miceblue425 36:50
2017-10-13Watts says Davina lets him bandwidth-limit a native 768 kHz recording to 48 kHz while keeping 768 kHz, to test whether bandwidth limiting and its filter type matter. miceblue425 40:47
2017-10-18Watts says inadequate bandwidth limiting harms sound via aliasing; Davina's first decimation filter has 260 dB rejection, and non-decimating FIR and IIR versions will let listeners hear bandwidth limiting. Head-Fi, Oct 2017
2017-12-02Watts says sampling theory only requires bandwidth limiting, which may use a non-pre-ringing IIR filter; his pre-ringing remark concerned DAC interpolation filters, and coefficient calculation matters beyond tap length. Head-Fi, Dec 2017
2018-01-03Watts says Hugo 2 filters DSD to 705.6 kHz PCM with an improved filter needing over 200 dB attenuation, so performance is limited only by the DSD itself. Head-Fi, Jan 2018
2018-01-03Watts says Hugo 1's DSD filter had audible aliasing at -150dB, the new one is better than -240dB, and this work informs Davina, which will make aliasing negligible. Head-Fi, Jan 2018
2018-01-16Watts says a sinc only perfectly reproduces signals with zero output at FS/2, but modern ADC half-band filters leave -6dB, causing unrepairable aliasing; Davina will reach -240dB. Head-Fi, Jan 2018
2018-01-16Watts says filtering out aliased content does not help as the damage is done, and longer tap length benefits all recordings, from 1990s passive anti-aliasing to modern half-band ADCs. Head-Fi, Jan 2018
2018-01-25Watts says 44.1 will be supported: the intent is to record at 705 kHz and decimate later, with uniquely aliasing-free decimation from 104.25 MHz down to 44.1 kHz. Head-Fi, Jan 2018
2018-01-25Watts says BBC Radio 3 on iPlayer, with two SRC stages (44.1>48>44.1) and AAC 320, is far below 44.1/16 and should not be called HD. Head-Fi, Jan 2018
2018-02-24Watts says SRC and half-band decimation damage timing, triangular dither degrades depth, so he buys 96k originals; Davina's filters reach about -240 dB and will test bandlimiting and dither. Head-Fi, Feb 2018
2018-02-26Watts says he has eliminated decimation aliasing, interpolation needs a closer-to-sinc filter, bandlimiting is likely a red herring, and he worries most about 16-bit reduction harming depth. Head-Fi, Feb 2018
2018-06-09Watts says lower sample rates bother him less than sample rate conversion, which adds distortion and noise and degrades timing, so buy a 96k original over its 44.1 version. Head-Fi, Jun 2018
2018-06-27Watts says aliasing ruins music: Hugo's -130dB DSD decimation was audible, Hugo 2, TT2 and M scaler reach -220dB, MQA manages -40dB, and 768 kHz needs 24 bits. Head-Fi, Jun 2018
2018-06-30Watts explains decimation needs low-pass filtering to avoid aliasing; decimating DSD to 705/768 kHz saves FPGA space, and Hugo 2's filter beats non-decimating DSD+ through better RF removal. Head-Fi, Jun 2018
2018-07-04Watts says Davina's 705/768 to 44.1/48 decimation filters tested perfectly with no aliasing; 104 MHz to 705/768 decimation design comes next, then hardware testing. Head-Fi, Jul 2018
2018-07-18Watts recounts WTA development from 1980s sinc ideals, finding sinc closeness beat image rejection alone, and wonders whether Davina's alias-free decimation will change the importance of tap length. Head-Fi, Jul 2018
2018-07-19Watts says modern ADCs' -6dB at FS/2 causes unrepairable aliasing, yet all recordings benefit from an M scaler; Davina will decimate without aliasing, and he will release randomly named files. Head-Fi, Jul 2018
2018-07-19Watts says it will identify losses from ideal 768 decimation then M scaling back up; the M scaler exceeds 16-bit recovery, but whether that suffices is open. needs context Head-Fi, Jul 2018
2018-08-04Watts says TT2 is warmer with better depth than Hugo 2 with the M Scaler too, and integer decimation from 192 to 96 loses little with a decent filter. Head-Fi, Aug 2018
2018-08-10Watts says Davina is on the back burner; 768 to 48 kHz decimation is done, a non-integer 104 MHz decimation filter exceeds 250 dB, and recordings are targeted next year. Head-Fi, Aug 2018
2018-08-16Watts says the new DSD decimation filter sounds slightly better than DSD+ through better RF suppression; Davina work showed decimating to 16FS is lossless if aliasing stays below -220 dB. Head-Fi, Aug 2018
2018-09-01Watts says reconstruction would match the analog signal to 16 bits in principle, but ADC aliasing from poor decimation cannot be repaired; he expects Davina to manage it. needs context Head-Fi, Sep 2018
2018-09-01Watts explains that 16-bit accuracy refers to the ideal sinc, errors arise only with transients, NOS is crude sample-and-hold, and delta-sigma ADCs suffer half-band decimation aliasing. Head-Fi, Sep 2018
2018-09-26Watts says he designed IIR and FIR Davina decimation filters with over 250 dB rejection to test bandlimiting; he expects pre-ringing at 22.05 kHz to prove a red herring. Head-Fi, Sep 2018
2018-09-26Watts shares Verilog simulation plots of his 768 kHz decimation filter with aliasing below -250 dB and 250 dB rejection at FS/2, versus -6 dB for half-band filters. Head-Fi, Sep 2018
2018-09-27Watts advises buying the original master, especially 96k, since SRC to 44.1 harms sound, but says 44.1 M scaled to 705k beats an unscaled 96k source. Head-Fi, Sep 2018
2018-10-23Watts explains the M scaler matches sinc to 16.7 bits, fixing transient timing, but cannot repair aliasing from modern half-band ADCs, possibly why 1960s recordings sound amazing. Head-Fi, Oct 2018
2018-11-07Watts says Hugo 2's DSD filter is his best, better than DAVE's DSD+ mode and any DSD-to-PCM utility, running at 705.6 kHz with more than 220dB filtering. Head-Fi, Nov 2018
2018-11-07Duplicate post: Watts says Hugo 2's DSD filter is his best, better than DAVE's DSD+ mode, running at 705.6 kHz with more than 220dB filtering. Head-Fi, Nov 2018
2019-02-05Watts says the prototype's 104MHz to 768kHz decimation filter failed in simulation, so he must redesign it from scratch, with little time due to other projects. Head-Fi, Feb 2019
2019-02-06Watts says true DSD has no decimation filter, conventional decimation gives about 20 dB rejection, and his FIR decimation is below -300dB; fractional 705/768 decimation was the problem. Head-Fi, Feb 2019
2019-02-09Watts prefers the M scaler's DSD filter, which decimates to 16FS with better than -220 dB above 88.2 kHz, to DAVE's DSD+, concluding aliasing causes decimation's sound problems. Head-Fi, Feb 2019
2019-02-26Watts gives DAVE's BNC channel mapping for the M scaler, says the DX output is for future products, and warns 705 to 768 kHz conversion adds distortion and noise. Head-Fi, Feb 2019
2019-03-25Watts comments on upsampling experiments: stopband beyond -150 dB matters little, output rate should suit tap length, his Davina decimation reaches 250 dB, and truncation needs aggressive noise shaping. Head-Fi, Mar 2019
2019-04-11Watts compares his FIR and IIR Davina decimation filters on a square wave, explains phase linearity, says both reach 250 dB rejection, and expects FIR to sound better. Head-Fi, Apr 2019
2019-04-25Watts says the M scaler's 56-bit output is reduced to 24 bits by -301 dB noise shapers; triangular dither sounds worse, and non-integer SRC degrades sound. Head-Fi, Apr 2019
2019-05-19Watts says DSD-to-PCM is not lossless; his Hugo 2, TT2 and M scaler filters exceed 200dB rejection with -301 dB truncation, against about 60 dB in chips. Head-Fi, May 2019
2019-05-25Watts explains decimation and aliasing, says tiny aliasing is audible, and preferred Hugo 2's -220 dB DSD filter to Hugo 1's -130 dB and DAVE's DSD+. Head-Fi, May 2019
2019-06-15Watts says 96k beats 48k mainly because cheap half-band ADC decimation filters alias, that SRC to 44.1k adds distortion, and that only signal-dependent timing errors matter. Head-Fi, Jun 2019
2020-06-04Watts says modern ADCs suffer aliasing from half-band decimation filters, noise floor modulation and poor small-signal accuracy, which his pulse array ADC will solve; the next prototype PCB is underway. Head-Fi, Jun 2020
2020-09-11Watts advises buying the original-rate release, such as 96/24 over 44.1/16, since sample-rate conversion adds timing errors, and says 24-bit sounds better than 16-bit; he does not stream. Head-Fi, Sep 2020
2020-10-29Watts says the M scaler and TT2 filter DSD, removing out-of-band noise with 220 dB attenuation and decimating to 705 kHz, which sounds better than DAVE's non-decimating DSD filters. Head-Fi, Oct 2020
2021-01-29Watts says his ADC uses a pulse array DAC in its feedback loop at 104 MHz, and converting to 768k synced to studio clocks needs a complex fractional interpolation filter. Head-Fi, Jan 2021
2021-05-04Watts says audible changes need a science-based explanation, citing ADC-project sample rate converter tests (aliasing from -200 to below -350 dB) his son heard blind, and M scaler tempo perception. Head-Fi, May 2021
2021-05-08Watts says he plans headphone tests of depth, impact and timbre for his ADC, microphones and headphones, plus a 768k-to-48k-to-768k test of decimation and WTA filters. Head-Fi, May 2021
2021-05-08Watts says content above half FS must be removed by decimation filters to avoid aliasing, and he must redo his -240 dB filters since -350 dB aliasing proved audible. Head-Fi, May 2021
2021-07-19Watts says ADC interpolator tests showed noise floor modulation audible near -360 dB, explains M scaler ferrites and Wave cables' solid ferrites, and suggests battery power plus optical for isolation. Head-Fi, Jul 2021
2022-04-25Watts says TT2 need only be separated if it thermally shuts down, explains how Blumlein encodes lateral and depth information, and says conventional SRC has severe sound-quality problems. Head-Fi, Apr 2022
2022-05-09Watts says decimation filters need only adequate aliasing rejection and FIR pre-ringing was inaudible, but DAC interpolation needs a sinc, so he offers one WTA1 filter, no options. Head-Fi, May 2022
2022-05-10Watts says decimation needs only removal of everything above the new Nyquist, so IIR works, but interpolation creates new samples and only a sinc reconstructs them perfectly. Head-Fi, May 2022
2022-07-12Watts says 16/44 gains more from M scaling but 24/192 M scaled is closer to the original; his goal is 24/48 matching 24/192, needing transparent 24-bit re-quantization at 48 kHz. Head-Fi, Jul 2022
2022-07-15Watts says removing pre-ringing harms transient reconstruction, impulses are illegal non-bandlimited signals, and his test filtering 20-22.05 kHz ADC aliasing after WTA1 gave little or no benefit. Head-Fi, Jul 2022
2022-08-19Watts says ADC decimation from 768 to 48 kHz is poorly done with inadequate anti-aliasing filters, and the planned test will measure both decimation and interpolation losses. AVstore 13:13
2022-08-19Watts attributes the blurred, soft transients to the ADC's decimation and inadequate anti-aliasing filters plus insufficient noise shaper accuracy, and says microphone siting was easier than expected. AVstore 48:06
2022-09-18Watts says his recordings test the playback chain and his pulse array ADC (begun about 2005); ADCs suffer noise floor modulation, poor small-signal accuracy suppressing depth, and sharpness-dulling aliasing. The Source AV 20:30
2023-03-16Watts says DAVE's DSD+ mode filters without decimating, yet the Hugo M Scaler's DSD conversion sounds better because its filtering is more effective, and explains decimation. Head-Fi, Mar 2023
2023-03-17Watts advises buying the original-rate release since WTA cannot undo non-integer SRC, and says the 2A charger advice is about holding 5 V at Mojo 2's port. Head-Fi, Mar 2023
2023-08-24Watts says Arrival at 48k 24b on Qobuz sounded far better than his 16b 44.1k copy, blaming non-integer sample rate conversion common on soundtracks. Head-Fi, Aug 2023
2023-08-25Watts says early Decca recordings have speed he attributes to aliasing and timing, suspects modern half-band ADC decimation, and says early PGGB sounded poorer than the Hugo M Scaler. Head-Fi, Aug 2023
2023-08-25Watts clarifies he means pre-1968 analog Decca tape, likely mastered via the Sony PCM1630 with passive anti-aliasing, versus today's half-band decimation at -6dB at 22.05kHz. Head-Fi, Aug 2023
2024-01-27Watts says the claim that SRC is inaudible is wrong, advises sending bit-perfect data without SRC or upsampling to his DACs, and says transient timing is vital for perception. Head-Fi, Jan 2024
2024-03-18Watts says Quartet's single BNC and optical outputs are fixed at 176.4/192k, decimated from 705.6/768 kHz with new filters and separately noise shaped for 4FS. Head-Fi, Mar 2024
2024-05-04Watts says Quartet replacing a Hugo M Scaler on another listener's Dave made instruments sound real, and describes ADC decimation filter work and the 104MHz pulse array ADC. Head-Fi, May 2024
2024-07-01Watts lists DAVE's three DSD options with a Hugo M Scaler; the scaler's DSD filter (better than 200dB rejection) sounds best, and he studies decimation limits via the ADC project. Head-Fi, Jul 2024
2026-10-11Watts says an ADC has a DAC in its feedback loop; the surprise was far more accurate pitch of female vocalists than pro ADCs give, which he attributes to aliasing. Distorsiuni Interioare 12:51
2026-10-11Watts says aliasing errors are not harmonically related to the signal, confuse the brain's pitch tracking and degrade transient timing; Quartet eliminates aliasing unlike standard pro ADCs. Distorsiuni Interioare 13:35
2026-10-11Watts says the ADC sounds better even on a phone because of decimation: conventional ADCs turn four or 5 bits at 3 MHz into twenty-four bits at 192 kHz. Distorsiuni Interioare 53:53
2026-10-11Watts says decimation uses cheap half-band filters only -6 dB at Fs/2, so a signal at 24 kHz plus 10 Hz reflects to 24 kHz minus 10 Hz. Distorsiuni Interioare 54:40
2026-10-11Watts says reflections continue down to DC at around -120 dB, with most aliasing between 18 and 24 kHz, which chip designers dismiss as inaudible. Distorsiuni Interioare 55:24
2026-10-11Watts says aliasing matters because out-of-band signal-correlated noise degrades transient timing, and Quartet's DSD decimation follows its ADC decimation process. Distorsiuni Interioare 56:10
2026-10-11Watts says the stronger decimation makes DSD converted to PCM sound better; going from 64FS to 16FS, DSD's high-frequency distortion must be removed. Distorsiuni Interioare 57:07
2026-10-14Watts says the pulse array ADC samples at 104 MHz, 5 bits, decimating to 48 or 768 kHz at 24 or 32 bits; decimation filtering is critical to transient reconstruction. Part-Time Audiophile 56:23
2026-10-14Watts says conventional decimation filters are only -6 dB at Fs/2, giving aliasing between 18 and 24 kHz at 48 kHz, with products around -120 dB in the best ADCs. Part-Time Audiophile 57:21
2026-10-14Watts says his ADC's aliasing is vanishingly small; comparing standard, high-end and pulse array ADCs on a female vocal from Daniel's simple analog tape, pitch was the surprise. Part-Time Audiophile 57:48
2026-10-14Watts says Quartet solved interpolation filters for him; with the ADC and 768 kHz recordings he can decimate (to '24 kHz' as spoken), re-interpolate and compare, controlling microphone to transducer. Part-Time Audiophile 1:03:53
2026-10-14Watts says tests comparing decimation and interpolation, and minimizing the difference, are proof that things are being done correctly. Part-Time Audiophile 1:04:48
2026-10-17Watts says Quartet returns lots, but the key question is the loss from decimating 768k to 48k and back, which needs a bank of test recordings through 2027. article, Oct 2026
2026-10-17Watts says 768 kHz Pulse Array ADC recordings can be decimated to 48 kHz and brought back to 768 kHz to measure losses, improving decimation and interpolation filters. article, Oct 2026
2026-10-17Watts says Quartet's 18 kHz filter in ADC decimation removes everything above 20 kHz, and that out-of-band noise and distortion are audible because they corrupt transient timing. article, Oct 2026
2026-10-17Watts says removing out-of-band noise outweighs the transparency losses of DAVE and Quartet, and an analog filter cannot do it transparently, so analog should stay simple. article, Oct 2026
2026-10-17Watts says wide-bandwidth amplifiers just let in rubbish and distortion, and filtering the high frequencies is better if done completely transparently, which digital can do. article, Oct 2026
2026-10-17Watts says two high-frequency filters in cascade sounded no different from one, which is how he knows they are transparent. article, Oct 2026