Rob Watts' 2017 RMAF video presentation
Rob Watts' 2017 RMAF video presentation is here: RMAF17: DAC Design Masterclass.
I think the slides are posted on Head-Fi.org.
Anyway, a main point Watts makes is that as taps increase, so do certain metrics (limits of Audio Precision test instrument) -- and that these are audible (e.g., -170 db noise )
Rob Watts' 2017 RMAF video presentation is here: RMAF17: DAC Design Masterclass.
I think the slides are posted on Head-Fi.org.
Anyway, a main point Watts makes is that as taps increase, so do certain metrics (limits of Audio Precision test instrument) -- and that these are audible (e.g., -170 db noise )
Last edited:
There's no argument that a longer filter can perform better.
As far as his claims go - of course he would say that at an event like RMAF. His story and claim are dubious at best.
It's a shame, IMO, because he's got a respectable product.
When you've got an FPGA hammer, everything looks like a nail I suppose. I guess I can't fault him too much, because at least his products objectively perform well.
Last edited:
Interesting link! When you select Pro Tools 2018 (or any other version of Pro Tools) and Transition, you see something that looks like the response of a rather mediocre halfband filter. The stopband suppression is about 75 dB from 23.1 kHz onwards and above 50 dB from around 22.9 kHz. So from 21 kHz...21.2 kHz onwards, there will be substantial aliasing in the audio signal. The passband is nearly flat until just above 21 kHz (the end of the passband and the start of the stopband always add up to the sample rate in a halfband filter).
Still, this shows that it could be useful to have a DAC filter passband that's a bit wider than the traditional 20 kHz at 44.1 kHz sample rate.
It's peculiar that minimum-phase filters have become so popular. In the early 1980's, the phase linearity of Philips CD players was their main selling point. Most competitors didn't use oversampling and had minimum-phase rather than phase-linear filters.
Yeah, many are better than the Pro Tools filters in that list. I know Ableton is quite popular and those filters look pretty good. I have no idea how accurate or comprehensive that link is, though. It's possible there are settings and variants that aren't listed.
why spartan6
As long as I know, Chord electronics in Japan announces that Dave has 166 DSPs and 164k taps FIR at 256OSR by employing Xc6slx75. Technically speaking, it's a little bit strange. Xc6slx75 has 132 DSPs and 172 block rams(172x18kbit). 164k taps FIR must have 164k coefficients. It means the coefficients is 18bit at max, though a logic cell is also configured to block memory. I would say the number of DSPs is a typo. But the capacity of coefficients is fatal. I can't understand why Chord doesn't use a large FPGA if they want looong tap FIR. There are many FPGAs which are inexpensive and have many resources than Xc6slx75. Xc7a100(artix7) is almost the same price but has more powerful 240DSPs and 270 block rams.
As long as I know, Chord electronics in Japan announces that Dave has 166 DSPs and 164k taps FIR at 256OSR by employing Xc6slx75. Technically speaking, it's a little bit strange. Xc6slx75 has 132 DSPs and 172 block rams(172x18kbit). 164k taps FIR must have 164k coefficients. It means the coefficients is 18bit at max, though a logic cell is also configured to block memory. I would say the number of DSPs is a typo. But the capacity of coefficients is fatal. I can't understand why Chord doesn't use a large FPGA if they want looong tap FIR. There are many FPGAs which are inexpensive and have many resources than Xc6slx75. Xc7a100(artix7) is almost the same price but has more powerful 240DSPs and 270 block rams.
By the way, I haven't got access to their paper at the moment, but Lagadec and Stockham published an interesting relation between linear-phase filter passband ripples and pre-echoes in a 1984 AES preprint (number 2097). Seemingly quite small ripples correspond to substantial pre-echoes. Since it relates to the passband, it has nothing to do with audible ultrasonics.
Yes, Bruno Putzeys and probably a few others have posted about that before I think. From what I understand, the pre-echo is an artifact of Equiripple filters designed via Parks-McClellan. Not to be confused with the totally expected pre-ringing of a linear phase FIR filter.
You need something to differentiate your product and attribute your competitors' "digital sound" to.
For some, the jury is still out on that.
There was a guy on another forum that was experimenting with, I believe, 5 million using Sox and he did say he could discriminate well over 1 million.
From the brief periods I had time to look in, it appeared an interesting and productive discussion.... not the sort of thing you see here very often these days.
Poor bloke contracted Meniere's disease and basically had to stop all his research.
T
Indeed, this is easy to test with a PC and Foobar ABX.
I have doubts about the validity of those results. I have known more than one audiophile that was unknowingly passing their entire chain through Windows 48 kHz resampling unknowingly for more than a year until I pointed it out. This is likely not a million tap audiophile approved filter given the minimal delay through it.
How would you reconcile this with the large movement to short delay slow rolloff or minimum phase filters? Surely, the 5 million tap FIR filters are exactly what they and the NOS folk are railing against. The problem here is I see zero consistency in preference, and a ton of award winning products using bog standard 256-tap or whatever filter is built into your AKM chip.
I'm all for the best technical solution, but I'm pretty sure the point of diminishing returns sets in somewhere well before the number of taps in your filter hits 7 figures. There are probably more important parameters in the interpolation filter to take care than throwing a million taps at whatever remez spits out. Anyway, I'm not claiming to be a DSP expert; I just don't think this passes the smell test.
Last edited:
Lagadec and/or Stockham had worked on a controlled listening experiment involving a filter bank. The filter bank had a linear phase response and a +/- 0.5 dB ripple all the way from 0 Hz to the Nyquist frequency, with 25 Hz distance between ripples (if I remember it correctly). They expected this to be barely audible, but in fact it was very audible and annoying to all listeners. They started to understand this when they looked at the time response: pre-echo, main peak and post echo with 40 ms distance between them and with the pre- and post-echo only some 30 dB down.
In general, when the main peak has a value of 1 and the pre- and post-echo have a value of alpha, the ripples in the steady-state sinewave response will go from 1 - 2 alpha to 1 + 2 alpha. When you don't want to rely on masking and want the echoes to be down by 120 dB (alpha = 1E-6), you can allow a passband ripple of +/- 0.000017 dB.
The whole story only holds exactly when the frequency distance between ripples is constant, which it usually isn't, but it is a reasonable estimate for typical Parks-McClellan filters. I checked it once by (mathematically) cascading a Parks-McClellan filter with an ideal filter and sampling the impulse response in what should be the zero crossings and found abberations that were within a few dB from the Lagadec/Stockham estimate.
I forgot to add that an equiripple filter is treated as a cascade of a hypothetical filter without passband ripples and a filter that only generates ripples. The Lagadec/Stockham theory is then applied to the latter filter. Hence, the calculated pre- and post-echoes are an artifact of the passband ripples, indeed quite apart from the pre- and post-ringing necessary to get filtering.
Good point, which leads right into that issue that some people seem to confuse pre-ringing (as is often displayed graphically in dac review articles) from pre-echo.
On another forum (head-fi ??), Watts was queried about some of those claims. He honestly noted that some came from software simulation (because no real-world test instrument, like AP, can measure distortion/noise THAT low).
Anyway, I don't know how the M2 Scaler measures, but it seems that DAVE's metrics have been superseded by the latest Benchmark:
Benchmark DAC3 HGC D/A preamplifier-headphone amplifier Measurements | Stereophile.com
Chord DAVE (a few mos earlier)":
Chord Electronics DAVE D/A processor Measurements | Stereophile.com
Last edited:
Minimal phase FIR is nothing different from linear phase FIR when it comes to tap length. Longer sounds better. I actually prefer IIR to short tap FIR, and I use minimal phase FIR. FIR sounds good only when the tap length is long enough for its filter frequency. Window also affect the sound quality a little. I don't know how those people have come to the conclusion that short FIR is better. I guess they are comparing extremely short ones, like 300 vs 6000. They both sound fine until you hear very long ones. You should test it yourself if you're interested in this subject, and I hope you'll understand what I'm trying to say...
Last edited:
You’re certainly welcome to your opinion. I don’t see any evidence or consensus supporting it though. I can design you a long filter that isn’t very optimized. Talking about the number of taps your filter has is backwards.
I’ve heard and generated extremely long filters done in software. I don’t agree.
BTW, I don’t think there are very many filters in commercial audio ICs that are windowed sinc. Maybe some audiophile parts using external custom filters. Everything else’s linear filters seem to be halfband equiripple.
Could be wrong, Marcel is the only DSP pro in this thread.
I’ve heard and generated extremely long filters done in software. I don’t agree.
BTW, I don’t think there are very many filters in commercial audio ICs that are windowed sinc. Maybe some audiophile parts using external custom filters. Everything else’s linear filters seem to be halfband equiripple.
Could be wrong, Marcel is the only DSP pro in this thread.
Last edited:
I'm curious how you feel the difference between long and short one. In my experience, it takes very long time (years) to recognize clearly the drawbacks and strength of each digital filter processing method. Even NEVE and SSL people were selling horrible sounding digital products for years without hesitation.
- Status
- This old topic is closed. If you want to reopen this topic, contact a moderator using the "Report Post" button.