The best DAC for under $10k is only $2k

Audio Gear Talk HQ Player
41

I use primarily four DACs, all “NOS” or I would rather call “bit-perfect” when fed with DSD data. Data from HQPlayer thus reaching the actual D/A conversion stage bit-perfect.

Primary DACs are T+A DAC 200, T+A HA 200, Holo Spring 2 and Holo Spring 3. I run all these always at DSD256 or DSD512 from HQPlayer.

I also have some other DACs in use that do the same at DSD. Such as iFi iDSD NEO and xDSD Gryphon (TI/BB chip that is always bit-perfect with DSD). In addition RME ADI-2 Pro, TEAC NT-503, Topping E30 and SMSL D-6 (all with AKM chip in DSD Direct mode). And then SMSL D300 (Rohm chip that is always bit-perfect with DSD).

Of these, Holo Audio is always bit-perfect with PCM data too. It doesn’t have any digital filters. I can run it at 1.4112/1.536 MHz sampling rate PCM data from HQPlayer. But I don’t use it that way much.

Not sure which filter where you are talking about here?

Most of the same filters are available for both PCM → PCM and PCM → SDM (DSD) output cases in HQPlayer. Difference is more like if you run upsampling to 32x PCM rate, the filter output is just 32x rate. While if you run upsampling to DSD512 rate, filter output is 512x rate.

2 Likes
42

Just curious. My understanding is that a digital filter is not needed if there is no sample rate change (and no DSP required), however, there obviously needs to be an analog filter at the output. From what I can tell the May NOS 44.1KHz analog anti-aliasing filter is very slow. Does that make it only a suitable band pass filter when fed high rate DSD (noise being pushed up to very high frequencies)?

43

Holo DACs have steeper analog filter than most chip based DACs.

By definition when May is NOS mode (Spring 3 is always NOS) there’s no digital filter and the DAC is bit-perfect. Thus there is only analog filter. I would never run such at low PCM rates.

The analog filter is sufficient to largely remove images when the converter is running at 32x PCM rates.

Since the DSD D/A conversion section is analog filter in itself too, the combination gives practically flat noise floor at DSD512 and DSD1024.

(ES9038PRO’s modulator noise peaks at around 700 kHz and digital filter output images around 352.8 kHz when running from 44.1/48k sample rate sources, depends on the DAC’s analog filter how much of this is left)

2 Likes
44

Just as Jussi mentions above, as high rate DSD as my computer can do is my preference. 1.5MHz PCM with LNS15 is also pretty good to my ears. I can’t do greater than DSD256 yet. Soon!

45

Hi Jussi,

Do you know if the Topping D30 Pro is any good? It uses the Cirrus CS43198 chip x4. I have one for use with my TV and it sounds pretty good, but I have not tried it critically yet. Thanks for all of your generous help in the fora over the years!

Rush

46

I have one of those too. It works OK, but I’m not sure if it ever uses the DSD Direct mode of the chip. Based on the chip datasheet, DSD128 is maximum you can get through the DSD Direct. DSD256 is supported through the non-direct DSD Processor path.

It works decently overall. If you have one, you can run it at DSD128 or DSD256, both will give you very similar performance and lower distortion than you can get with any PCM input rate. If you need a device that gives you decent performance already at DSD128, this is one of those.

1 Like
47

Thank you Jussi! Great info!

48

Thanks for confirming that the Holo May DAC needs an upsampler to play NOS correctly without aliasing from the reconstruction filter. This is what I understood from the Stereophile review.

I just find the design a bit puzzling. A stair step NOS output is obviously loaded with ultrasonic aliased frequencies. The goal seems to be a near perfect impulse response - something that might interest only a bat or any creature able to hear extreme ultrasonics - not sure I can fathom what purpose this capability serves but the problems it creates are real.

49

That’s what you get with PCM in general, since that’s what PCM is about. Stair steps only disappear once you reach high enough oversampling rate combined with sufficient analog filter. If you don’t want to spoil the phase response, you need to have the analog filter corner at 100 kHz or higher. Which means that your digital filter oversampling rate needs to be in MHz range.

But the problem also plagues SDM chip DACs because they don’t have enough DSP power to run proper digital filters with such ratios. ESS runs mediocre quality digital filters up to 8x the input rate. As result, the output is still stair-stepped with less than 16-bit output accuracy.

Just right now, I’m listening internet radio with HQPlayer, running digital filters to 512x input rate and >300 dB attenuation, followed by my best delta-sigma modulators. Output is through Holo Audio Spring 3 at the moment at DSD512. (yeah, source is just bare 320 kbps MP3 stream :joy: but why not)

Plus some of my clean-up-the-MP3-junk-DSP.

CPU load on Ryzen 7 5800X is not too bad though:

Screenshot from 2023-01-15 04-52-46
Screenshot from 2023-01-15 04-52-46786×592 122 KB

3 Likes
50

I don’t quite agree that things are quite so black & white. More like some merits here some compromises there etc.

I don’t regard pre-ringing as a problem at all - a complete blind alley more like - it is the perfect result to be expected from a sharp linear phase filter when hit with an impulse and is totally accurate. No musical signals will “pre-ring” in the slightest as they must all be well below the frequency of the brick wall. Pre-ringing is entirely a red herring that has caused much confusion in the community.

PCM is by far a better format than 1 bit DSD. Practically speaking no commercial music is recorded direct to DSD (so few recordings as to be negligible).

PCM in general is not outputting stair steps nor are ESS chips (most DAC have an appropriate analog reconstruction filter). Holo May KTE does stair steps in NOS mode due to lack of a reconstruction filter. Thank you for confirming that upsampling to high rate DSD or high rate PCM helps to overcome this limitation of the Holo May analog reconstruction filter.

Destroying the phase response is the major problem caused by minimum phase filtering necessitating such a high corner frequency. Minimum phase filtering often creates easily audible problems. “Correcting” pre-ringing is not a correction - it actually adds phase distortion unless filter corner is sufficiently high enough not to affect in band audio at all - yet it is marketed as a correction even though inaudible for properly bandpass filtered audio.

Linear Phase filters preserve phase perfectly. Nothing new - engineers always chose them without hesitation 30 years ago - lately we seem to have forgotten physics.

Many independent measurements show multiple modern ESS DACs have 21 bit resolution. If your 16 bit accuracy is due to passband ripple then this isn’t really a concern since there would be passband ripple on the ADC anti-alias filters used to record the instruments in the very first place.

Very high rate DSD conversion fed into a DAC that natively handles DSD does eliminate the effect of poor PCM reconstruction filter choices by manufacturers available in many DACs (slow leaky and minimum phase filters).

If anything ESS made a mistake by offering many poor filter choices on their latest chips - only 1 of their 8 filter options is actually acceptable. I do feel ESS with the 1 proper filter works extremely well but agree with you that performance is poor with 7 out of 8 choices.

Your responses and insights are much appreciated. Your understanding is accurate. If anything it is the degree of relative importance of certain things that I can’t square with my understanding.

51

But they are not well below the frequency of brick wall. If your ADC rate is 44.1k the frequency limit is 22.05 kHz. But many instruments produce content way up to 60 kHz. And with percussions and such you can easily reach 100 kHz.

In any case, HQPlayer has running on the fly analysis of the source data to detect problematic source content and shows you incrementing counter value from the detector.

Pre-ringing is of course only one of the many problems baked into the source data.

I’m offering linear, intermediate and minimum -phase apodizing and non-apodizing filters. You can easily try and compare which way you prefer. I’m not trying to force anybody to use minimum-phase filters. Current default filter settings in HQPlayer are linear phase.

Pre-ringing is just unnatural and doesn’t happen with real world signals, because in real world there’s no sound before the physical event causing the sound has happened.

For some of the studio productions content I prefer minimum phase filters myself. For example with old Pink Floyd albums from early 70’s.

You haven’t looked at NativeDSD site yet? For example I personally know pretty well two persons who are doing commercial DSD recordings (Channel Classics).

BlueCoast for example mixes in analog. These days, a lot of both PCM and DSD recordings are actually mixed in analog. Using analog production gear has become very popular again.
Most famous analog desks you can buy from Rupert Neve: https://rupertneve.com
Another big name SSL is also making analog desks: Music Production & Mixing | Solid State Logic
And so is AMS-Neve: https://www.ams-neve.com/

Essentially you record with multichannel PCM or DSD ADC, then play it back through DAC to the mixing desk while recording the desk output again through PCM or DSD ADC. So exactly same work flow you would have with old school analog multitrack tapes.

I’m measuring these devices all the time, as yes they do.

It has better analog reconstruction filter than most of your ESS DACs.

You are looking at wrong measurements.

Here you can see 0 - 22.05 kHz sweep, peak hold, from Topping D90SE DAC. Peak image level is about -72 dB compared to the signal level giving 12-bit worth of reconstruction accuracy here. That image around 352.8 kHz is the left-over stair stepping. And this is after the analog output reconstruction filter.

Topping_D90SE-sweep-44k1-narrow
Topping_D90SE-sweep-44k1-narrow1280×720 8.78 KB

2 Likes
52

Ok I understand your view now. More emphasis again on accuracy (looking at pure numbers) where physics of sound indicates it just isn’t necessary or even useful.

Recorded audio should have been anti-aliased at 20KHz prior to AtoD. No 22 KHz signals should ever be there. I also don’t believe that inaudible frequencies are of any use in an audio reproduction system. I stick with the conventional view that 20KHz is the extreme upper end of human hearing (rarely encountered in adults). 20KHz is already overkill for music. Typical excellent hearing in adults being rarely above 17 KHz and declining with age. There is very little if anything musical above 12 KHz anyway.

53

Localization of sound occurs at high frequencies. There is an increasing body of evidence that, even though we don’t actively hear some frequencies in a started hearing test, we may still be using those frequencies to determine where a sound is coming from. Therefore, accurate reproduction of the high frequencies produced by instruments is important to capture and reproduce if one wants to accurately recreate the spatial cues of the original venue for which the recording occurred.

2 Likes
54

OK, lets go a bit deeper on this. First, you cannot do analog anti-alias cutting at 22 kHz to avoid aliasing at 44.1 kHz sampling rate while preserving flat phase response. Again, to preserve even nearly flat phase response, analog filter needs to have corner at 100 kHz or higher.

So the approach is to oversample the input with SDM (DSD-style) ADC and then in digital domain decimate this down to 44.1 kHz rate with a linear-phase brickwall anti-alias filter. That filter will produce ringing.

Because 44.1 kHz sampling rate leaves you 2050 Hz wide band to transition between 20 kHz and 22.05 kHz Nyquist frequency.

1 Like
55

Hearing is not doing like Fourier transform.

Instead it is more like wavelet transform matching different wavelets. It primarily detects wave fronts, essentially signal’s step response. Instead of steady state signals.

For musical instruments too, the attack phase is more important for the character than the sustained state or decay state parts.

From Fourier perspective, it means passing very high frequency content, because rapid waveform changes require high frequencies in Fourier transform. This is however different from hearing discrete constant tones of those same frequencies.

This is very complex for sparsely sampled low resolution digital sources like RedBook. Sample timing rarely coincides with the actual event, it always happens “somewhere between the samples”. Thus the reconstruction (digital and analog) filter is critical in mathematically assembling the correct transient structure from the surrounding data.

4 Likes
56

There are several sound location methods

  1. Between 200 and 2KHz it is mostly inter-aural arrival delay (using two ears)
  2. above about 6KHz there is a strong shadowing of sound from one ear to the other for sources off axis as little as 30 degrees can be as much as 10 db.
  3. Pinea shape and distortion according to source direction (why most headphones are a waste of time and make sound from inside your head)
  4. reverb clues or echoes - tell you about the type and size of space, height of a sound above a surface, walls etc.

There may be others but nobody has demonstrated that inaudible high frequencies do anything at all - you need to research this by conducting trials and publish papers - you will be famous.

2 Likes
57

https://journals.physiology.org/doi/full/10.1152/jn.2000.83.6.3548

Also from other perspective:

1 Like
58

You should research and demonstrate that humans hear high frequency sounds well beyond 20KHz and up into the 40KHz range (in between samples). You will be famous. Since nobody has demonstrated this so far, I prefer to stick to accepted science. This doesn’t mean attack isn’t terribly important - it is - but anything above 20 KHz will not be heard.

Again, I don’t disagree with you fully (attack and timbre are extremely important) - just the extrapolations to extremes aren’t validated (so far).

1 Like
59

It is not about hearing sine above 20 kHz, but hearing differences in attack waveform shapes with rise time details faster than what can be described by 50 microsecond units.

And you know how Dirac pulse will become through these digital systems.

But this is besides the point. More bandwidth (sample rate) limiting you do, more problems you have in time/frequency/phase. Just pick a system where you have least bandwidth limitation. RedBook is the worst, so crammed into a small box that it is the most challenging to make work out well in practice.

At DSD512 you already have 200 kHz worth of audio bandwidth and your Nyquist frequency is at 11.3 MHz. Hardly a limiting factor anymore.

2 Likes
60

As described this ringing will affect frequencies at the corner frequency of the brick wall anti-alias filter. Ringing will not be present with any audible musical signals. Signals up to 20 KHz may be well preserved with no ringing if the filter is well chosen.