I think there is no definite answer to that at the moment. Well-informed people have very different opinions about this very subject.
Controlling time smearing and pre-ringing is exactly what that article of Peter Craven is about. (By the way, please check your PM.)
After spending bit more time on reading and listening, it seems that with NOS DAC's, the FIR compensation filter implementation as executed by Totaldac and Schiit (closed form filter) could be the answer. I haven't listened to either of them.
I personally settled on NOS DAC with no filters of any kind, no oversampling, but with an increase in high-frequency spectrum done in JRiver DSP Studio Parametric Equalizer: 20kHz, Q=0.3, and Gain of 3dB. This produces the full sound presentation, very natural, with excellent definition of voices/instruments.
Even the slight increase in oversampling causes unlistenable results - smearing of definition. The initial transients are also pretty much completely gone by the time I oversample sufficiently enough to compensate for the NOS -3dB loss @ 20kHz.
Post-filtering causes pretty much the same effects. This time the high frequencies suffer more.
The PCM / DSD DAC sounds correct for the first few minutes of listening to it, but this correctness vanishes after that period and is replaced by a completely unnatural sound presentation that is fatiguing, to say the least. I can see how such a sound could be accepted as perfectly okay by many... until it is compared to a true NOS.
I personally settled on NOS DAC with no filters of any kind, no oversampling, but with an increase in high-frequency spectrum done in JRiver DSP Studio Parametric Equalizer: 20kHz, Q=0.3, and Gain of 3dB. This produces the full sound presentation, very natural, with excellent definition of voices/instruments.
Even the slight increase in oversampling causes unlistenable results - smearing of definition. The initial transients are also pretty much completely gone by the time I oversample sufficiently enough to compensate for the NOS -3dB loss @ 20kHz.
Post-filtering causes pretty much the same effects. This time the high frequencies suffer more.
The PCM / DSD DAC sounds correct for the first few minutes of listening to it, but this correctness vanishes after that period and is replaced by a completely unnatural sound presentation that is fatiguing, to say the least. I can see how such a sound could be accepted as perfectly okay by many... until it is compared to a true NOS.
You can employ a soft sloping (non-brickwall) reconstruction filter mode. Such soft filter modes are a very common feature among today's DAC's. They sacrifice a bit of upper band edge flatness to shorten impulse response ringing duration. They also allow some leak-through of the first image-band.
Soft sloping filters cannot, however, be employed for ADC anti-aliasing, as they would allow some in-band aliasing.
It's beneficial to look at the entire system ...
First, the analog-to-digital conversion: By its very nature, digital audio is a sampled-data system ... and in this system, like most sampled-data systems, aliasing is a time-variant distortion to be avoided at all costs. By far, the best method to avoid aliasing is a brick-wall, linear-phase, anti-alias filter (cutoff ~ Fs/2) prior to the sampling event. Three reasons why:
1. The brick-wall, linear-phase, anti-alias filter preserves all frequency-domain content (both amplitude and phase) of the signal to-be-sampled below Fs/2 (which is the only frequency range we can hope to capture, or preserve, in a sampled data system).
2. The brick-wall, linear-phase filter removes all content above Fs/2, that would otherwise alias into the frequency band below Fs/2 upon sampling.
3. If there is no content in the signal to-be-sampled above Fs/2, then the brick-wall, linear-phase, anti-alias filter has no impact on the signal at all ... in either the time or frequency domains.
And now, the digital-to-analog conversion (the point of this thread): Once we accept the need for anti-aliasing (on the ADC "front end") ... and the ideal filter to prevent aliasing ... then any subsequent brick-wall, linear-phase filters in the system at Fs/2 have no effect at all ... in either the time or frequency domains (since two or more brick-wall filters with the same cutoff in cascade are indistinguishable from a single filter).
The conclusion is rather unavoidable : unless there's some pleasing sonic consequences to the presence of post-DAC, ultra-sonic images ... which were definitely NOT present in the original signal ... there's no reason to avoid oversampling DACs with sharp, linear-phase, anti-image filters. It's a sharply band-limited system to begin with ... like it or not ... and no argument about post-DAC filter slopes can ever change this simple fact of life.
First, the analog-to-digital conversion: By its very nature, digital audio is a sampled-data system ... and in this system, like most sampled-data systems, aliasing is a time-variant distortion to be avoided at all costs. By far, the best method to avoid aliasing is a brick-wall, linear-phase, anti-alias filter (cutoff ~ Fs/2) prior to the sampling event. Three reasons why:
1. The brick-wall, linear-phase, anti-alias filter preserves all frequency-domain content (both amplitude and phase) of the signal to-be-sampled below Fs/2 (which is the only frequency range we can hope to capture, or preserve, in a sampled data system).
2. The brick-wall, linear-phase filter removes all content above Fs/2, that would otherwise alias into the frequency band below Fs/2 upon sampling.
3. If there is no content in the signal to-be-sampled above Fs/2, then the brick-wall, linear-phase, anti-alias filter has no impact on the signal at all ... in either the time or frequency domains.
And now, the digital-to-analog conversion (the point of this thread): Once we accept the need for anti-aliasing (on the ADC "front end") ... and the ideal filter to prevent aliasing ... then any subsequent brick-wall, linear-phase filters in the system at Fs/2 have no effect at all ... in either the time or frequency domains (since two or more brick-wall filters with the same cutoff in cascade are indistinguishable from a single filter).
The conclusion is rather unavoidable : unless there's some pleasing sonic consequences to the presence of post-DAC, ultra-sonic images ... which were definitely NOT present in the original signal ... there's no reason to avoid oversampling DACs with sharp, linear-phase, anti-image filters. It's a sharply band-limited system to begin with ... like it or not ... and no argument about post-DAC filter slopes can ever change this simple fact of life.
Last edited:
This does appear to me to assume that the DAC operating at the faster (oversampled) rate is of no lesser performance than the same DAC operating NOS. An assumption which no multibit DAC that I've ever encountered fulfills.
Yes, that's the logical conclusion. ADC anti-alias SINC filter idempotency should render DAC anti-image filters audibly inconsequential. Except that NOS DACs produce sound very obviously audibly different from that of Oversampled DACs, despite brickwall ADC band-limiting. Whether one prefers the sound of NOS to OS isn't the most interesting question, to me. The most interesting question is, why do the two produce so different a sound character at all? What audibly 'pleasing consequences' could possibly be produced by image bands which are totally ultrasonic? That's the primary technical mystery behind NOS DAC sound.
The first explanation that is usually suggested is that there is probably an system chain based intermodulation sensitivity to the unfiltered ultrasonic image products of NOS. However, the character of a NOS DAC remains fairly consistent across many differing systems, making it unlikely that system chain based intermodulation is the cause of NOS sound. Later suggested explanations tend to focus on time-domain arguments. Whatever the cause, it's yet to be adequately explained, as far as I know.
Last edited:
It seems that creating, on-a-PC, a digital brick wall filter which does nor pre and post ring could be of interest to filterless NOS DAC. This should be correct in a time domain, and decently flat in a frequency domain.
There is probably a misunderstanding about a band-limited system and impulse response. The attached is a simple example of a real impulse response. I want to emphasize "real" because I have seen several virtual ones which will not exist in real world.
L-channel (white) is 6kHz impulse waveform sampled by 192kHz. It roughly consists of DC,6kHz,12kHz,,, and 96kHz. Their amplitude is exactly same and phase-aligned. I need to use rough description. The band-limited system will not allow me to describe true situation. I don't discuss this because this is too complicated and not my intention here.
This expression, 6kHz impulse, may sound strange. I know there are many impulses which have no definition of frequency. I suppose they have an infinitely long period. In another word, the frequency is infinitely low(0.001Hz). This assumption is almost correct in many cases. But if you use impulse response in an audio application, you must define the frequency to prevent meaningless misunderstanding.
R-channel(blue) is the output of brick wall LPF(20kHz cut off) with linear phase. It means R-channel has only three sine waves,6kHz,12kHz, and 18kHz as you see in FFT plot. This is a normal output of x4OS DAC without post analog LPF. R-channel seems to have pre and post ringing, which are usually not preferred, from the waveform in time domain.
But FFT says they are three distortion free sine waves which are included in the input signal. So, the brick wall filter successfully did the job. Has the waveform distortion? FFT says no, definitely. It has ringing but not distortion. Ringing does not always mean distortion. The waveform in time domain sometimes leads you to misunderstand. You need to use another tool to know the real meaning. In this example, ringing means three distortion free sine waves. It has no other meaning.
L-channel (white) is 6kHz impulse waveform sampled by 192kHz. It roughly consists of DC,6kHz,12kHz,,, and 96kHz. Their amplitude is exactly same and phase-aligned. I need to use rough description. The band-limited system will not allow me to describe true situation. I don't discuss this because this is too complicated and not my intention here.
This expression, 6kHz impulse, may sound strange. I know there are many impulses which have no definition of frequency. I suppose they have an infinitely long period. In another word, the frequency is infinitely low(0.001Hz). This assumption is almost correct in many cases. But if you use impulse response in an audio application, you must define the frequency to prevent meaningless misunderstanding.
R-channel(blue) is the output of brick wall LPF(20kHz cut off) with linear phase. It means R-channel has only three sine waves,6kHz,12kHz, and 18kHz as you see in FFT plot. This is a normal output of x4OS DAC without post analog LPF. R-channel seems to have pre and post ringing, which are usually not preferred, from the waveform in time domain.
But FFT says they are three distortion free sine waves which are included in the input signal. So, the brick wall filter successfully did the job. Has the waveform distortion? FFT says no, definitely. It has ringing but not distortion. Ringing does not always mean distortion. The waveform in time domain sometimes leads you to misunderstand. You need to use another tool to know the real meaning. In this example, ringing means three distortion free sine waves. It has no other meaning.
Attachments
A non-ringing brickwall filter cannot be achieved. A brickwall filter MUST ring, and a non-ringing filter CAN'T have a brickwall frequency response. The mathematical proof of this is found in the Fourier transform of the two.
It's the Gibbs Phenomenon: an abrupt discontinuity in one domain (frequency or time) manifests as ringing in the other domain (time or frequency). No way to avoid it.
... and sampled-data systems (like digital audio ... at least, where the sampling frequency is close to 2x our hearing limit) MUST be abruptly discontinuous, in the frequency domain, if aliasing is to be avoided.
The system has already been abruptly, sharply bandlimited by the anti-alias filter prior to sampling. If there was spectral energy beyond Fs/2, it was removed ... sharply ... to prevent aliasing. If there was no spectral energy beyond Fs/2, the anti-alias filter had no effect (in frequency or time). In short, the ringing is already "baked-in" ... as it must be, in any sampled-data system ... long before any post-DAC filtering.
Sure, some may find some pleasing colorations from rolled-off treble, ultrasonic images or nonlinear phase distortion. But, at least in the ideal sense, there's only one way to preserve magnitude and phase accuracy below Fs/2.
If you allow some treble roll-off, you can make what Peter Craven calls an apodizing filter: a relatively smooth filter with relatively little pre-ringing that has a stopband that starts at a lower frequency than the transition bands of all brickwall filters in the signal chain. Assuming that all brickwall filters have linear phase and negligible passband ripple, the magnitude response and phase response (except for a constant delay) of the entire chain are then determined by the apodizing filter. The impulse response is simply the inverse Fourier transform of the response and is also determined by the apodizing filter (except for a constant delay). Hence, the apodizing filter determines the pre- and post-ringing.
Whether it makes sense to do this depends on the sample rate and on whether you believe in the audibility of supposedly ultrasonic ringing. For the CD sample rate it would mean a roll-off far below 20 kHz and no sound at all anymore at 20 kHz.
For what it's worth, the hypothesis that the thread starter's subjective preference is due to slight treble roll-off caused by the zeroth-order hold also sounds more likely to me than any theory about audible ultrasonics, but that's just my personal prejudice.
Whether it makes sense to do this depends on the sample rate and on whether you believe in the audibility of supposedly ultrasonic ringing. For the CD sample rate it would mean a roll-off far below 20 kHz and no sound at all anymore at 20 kHz.
For what it's worth, the hypothesis that the thread starter's subjective preference is due to slight treble roll-off caused by the zeroth-order hold also sounds more likely to me than any theory about audible ultrasonics, but that's just my personal prejudice.
About Linear Phase vs. Min. Phase vs. etc. reconstr./digital filters, the Stereophile/Keith Howard 2006 article -- Ringing False -- came to no definite conclusions as to which sounds better.
I think the test/comparison was a good one as each filter type, using ScopeDSP, was implemented in the same device.
Since then several mainstream manuf. (Meridian, Ayre, Cambridge; and DAC IC manufs, also) have claimed benefits for their proprietary "apodizing" or Min. Phase filters. (These are implemented in DSP/FPGA, with purported custom algos and coefficients).
I'm not sure about which DAC chip manuf. (ESS, AKM, TI, Cirrus) have their own firmware "apodizing" filters built in (if any).
That said, here are a few options from CS:
CS43130 (has several built-in, including non-oversampling)
WM8741 (Selectable advanced digital filter responses; Includes linear/minimum phase and range of tailored characteristics)
I think the test/comparison was a good one as each filter type, using ScopeDSP, was implemented in the same device.
Since then several mainstream manuf. (Meridian, Ayre, Cambridge; and DAC IC manufs, also) have claimed benefits for their proprietary "apodizing" or Min. Phase filters. (These are implemented in DSP/FPGA, with purported custom algos and coefficients).
I'm not sure about which DAC chip manuf. (ESS, AKM, TI, Cirrus) have their own firmware "apodizing" filters built in (if any).
That said, here are a few options from CS:
CS43130 (has several built-in, including non-oversampling)
WM8741 (Selectable advanced digital filter responses; Includes linear/minimum phase and range of tailored characteristics)
finally someone with real listening experience and not just crazy theories and scopes
AFAIK the latest offerings from ESS (ES9028 & ES9038) do and so do several chips from AKM (AK4490, AK4495, AK4497 come to mind). ESS also supports the implementation of custom filters by uploading custom coefficients, but the number of taps that are supported is not what you would call impressive.
- Status
- This old topic is closed. If you want to reopen this topic, contact a moderator using the "Report Post" button.