Which HQP Filter are you using? [2023]

[jussi_laako (Jussi Laako (Signalyst))]

There are various, with different weightings from the middle towards the extremes.

But I think the current default set is a good starting point. 1x=poly-sinc-gauss-long Nx=poly-sinc-gauss-hires-lp

If you want to go towards the longer extreme without overdoing, poly-sinc-gauss-xl(a). But in the manual there are some suggestions for different source content genres. Different music genres have different signal properties. But the default values are pretty universal and what I personally use most of the time.

[Mediahound]

Which dither do you use most of the time?

[jussi_laako (Jussi Laako (Signalyst))]

Most of the time I output SDM through ASDM7ECv2, so dither selection is not applicable.

[Mediahound]

Thanks. Which one do you recommend if playing PCM?

[jussi_laako (Jussi Laako (Signalyst))]

It depends on the output rate. R2R for lower rates like 8x, NS9. For higher rates like 16x+ LNS15 or NS5. NS9 can be used minimum 4x rates already.

If your DAC does further DSP (delta-sigma DACs), then TPDF or Gauss1 dither is usually just fine.

[Shawn_Williams]

It’s almost like no one reads the manual

[Michael_Weiland]

Any experience with a Mac Studio Ultra and DSD512 + gauss long + gauss hires lp?

[KML]

Hi @jussi_laako,

I have been enjoying HQPlayer for more than a year. In my humble opinion, it really adds a huge value to Roon, and makes music a lot more enjoyable, and I really want to thank you.

I have been comparing a lot of filters during the past months, and I would like to make sure that I understand some upsampling concepts correctly.

I apologize for that somewhat lengthy post, especially if the following questions have been answered before.

1. What is attenuation? For example, poly-sinc-gauss-XL has “extremely high attenuation”. Does it refer to the very slight decibel drop in higher frequencies (15-20Hz or more) that is often perceived as making the sound a little “darker”, “less harsh” or “bright” in the trebles?

2. Now, sinc-M has “very SHARP cut-off”. Does this expression refer to the very steep slope that falls almost perpendicularly past a certain frequency on a graphical representation or spectrum analysis? I understand that those very steep slopes are indifferently referred to as “fast” or “sharp” “roll off” or “cut-off” : are these terms equivalent?

So, if I understand correctly those concepts, sinc-M goes as following: the trebles (say past 15 Hz or more?) are “highly attenuated”; past a certain frequency, they fall abruptly, meaning that hypersounds are almost completely filtered. With poly-sinc-gauss-XL, the attenuation is even higher. (But how about the slope?)

3. What does the Lenght of a filter refer to? For example, the manual says that poly-sinc-ext-3 is 8 times longer than ext2. I think I read on this forum that the longer the filtering, the more effective it will be. But effective at what? Removing sound artifacts? Is that the proper meaning of the word?

4. What does apodizing mean? From my non-technical point of view, it means that it filters some of the artefacts in a recording. Is that definition satisfactory?

I am really trying to make sense of the big differences in sound that I clearly hear between my favourite filters in pcm: sinc-M, sinc-L, poly-sinc-ext3, and poly-sinc-gauss-Xla.

To my ears, sinc-L offers the most precise, natural and transparent sound : timbers, tonalities, textures, everything is presented correctly, but it also seem to let every artefact through, and has a fair amount of stridency on certain recordings, so it can be very fatiguing. Is that mainly because it has less attenuation that say, sinc-M or poly-sinc-gauss-Xla, or rather that its slope is extreme, or that is that it is not an apodizing filter, or something else I could be missing?

To my ears, sinc-M is, by comparison, a little softer, more veiled, but it is also more “dense”, “organic”, “substantial” : it shows the notes from the inside rather than sharply outlining them. It is also less fatiguing. I think I like it the most from a “musical” point of view (rather than an “audiophile” point of view), because it removes artifacts and any non-musically pertinent micro-details; it lets you focus on the melodic lines rather than being constantly distracted. Is that mainly because it has a higher attenuation and a less pronounced cut-off than sinc-L, or because it is apodizing, or is there something else in play?

(I’d also be curious to know how others perceive those filters!).

Thank you so much for reading this post.

[jussi_laako (Jussi Laako (Signalyst))]

@KML

What is attenuation? For example, poly-sinc-gauss-XL has “extremely high attenuation”. Does it refer to the very slight decibel drop in higher frequencies (15-20Hz or more) that is often perceived as making the sound a little “darker”, “less harsh” or “bright” in the trebles?

Attenuation, aka stop-band attenuation, is how much filter attenuates frequencies above cut-off (transition band).

@KML

Now, sinc-M has “very SHARP cut-off”. Does this expression refer to the very steep slope that falls almost perpendicularly past a certain frequency on a graphical representation or spectrum analysis? I understand that those very steep slopes are indifferently referred to as “fast” or “sharp” “roll off” or “cut-off” : are these terms equivalent?

Yes, that’s correct, and those terms are equivalent. It means the filter has very narrow transition band.

@KML

With poly-sinc-gauss-XL, the attenuation is even higher. (But how about the slope?)

It is also very steep slope. If we take 44.1k input rate as an example, transition band is about 250 Hz wide and reaches maximum attenuation at Nyquist frequency fs/2 at 22.05 kHz.

@KML

What does the Lenght of a filter refer to? For example, the manual says that poly-sinc-ext-3 is 8 times longer than ext2. I think I read on this forum that the longer the filtering, the more effective it will be. But effective at what? Removing sound artifacts? Is that the proper meaning of the word?

It is about length of the filter in time domain, in other word number of taps. If affects only width of the transition band. Longer the filter, steeper the slope. IOW, very slow roll-off filter is very short. It is not really any measure of effectiveness.

@KML

What does apodizing mean? From my non-technical point of view, it means that it filters some of the artefacts in a recording. Is that definition satisfactory?

It has two functions:

1. Any ADC anti-alias / decimation filter will leave it’s “fingerprint” on the data. Apodizing filter can modify this fingerprint.
2. Typical ADC anti-alias / decimation filter is pretty imperfect and leave errors to the data, these errors can be cleaned up with an apodizing filter.

In addition to ADC, these can also originate from the mastering software tools.

@KML

but it also seem to let every artefact through, and has a fair amount of stridency on certain recordings, so it can be very fatiguing. Is that mainly because it has less attenuation that say, sinc-M or poly-sinc-gauss-Xla, or rather that its slope is extreme, or that is that it is not an apodizing filter, or something else I could be missing?

Typically that is because it is non-apodizing, and secondarily for the other reasons.

@KML

Is that mainly because it has a higher attenuation and a less pronounced cut-off than sinc-L, or because it is apodizing, or is there something else in play?

Both have such similar cut-off steepness, that it is unlikely to be main difference. Primarily likely the apodizing feature. But also due to plain different filter type.

[Michael_Weiland]

@Michael_Weiland

Any experience with a Mac Studio Ultra and DSD512 + gauss long + gauss hires lp?

No one? Can no one confirm this?

[Shawn_Williams]

There is a discussion about it on another forum starting here

HQplayer 4 and Mac-mini M1

[KML]

Thanks for the graphic, @jussi_laako, it helps. If I understand correctly, a slow roll-off (or cutoff) in the transition band is one of the main characteristics of a short filter (fewer taps in time-domain), and necessarily has a gradual slope in frequency-domain (that can even start to roll-off in the passband?)

Conversely, a fast filter in time-domain (also referred to as “sharp” in frequency-domain?) always has a steeper slope, and a longer transition band (with more taps). In other words, length or speed (of a given filter in time-domain), and steepness (of the cutoff frequency) are always related. I am putting it together correctly?

I am trying to make sense of what I hear. Suppose I use a slow/super slow filter : I clearly perceive that the trebles are somewhat attenuated. It often gives me the impression that the sound is a bit less transparent, somewhat less “sharp” or strident in the highs. Is that because a slow cutoff in the transition band is inevitably correlated to some attenuation in the passband (in the audible range, say around 15 kHz or higher, and up)?

I guess my main question is : how are the attenuation of the passband and that of the stopband related? Does a super slow filter necessarily have a lower cutoff frequency (thus rolling off some of the trebles in the passband?)

From what I see in a typical representation of digital filters, a steep slope is most of the time correlated to a higher frequency of the passband and to a higher cutoff frequency. Is that correct? Conversely, I see slow filters start to roll-off trebles in the passband. In plain words, when I hear more trebles and transparency, is it mainly because of more stopband attenuation, or because stopband attenuation inevitably means more passband attenuation? Am I completely off base?

Thanks again for helping us understand those technical terms.

[MistaLovaLova]

I really like your description of the filters, they sound very similar to how I perceive them. Sinc-L sounds the most “vivid”, punchy and dynamic, although sometimes at the cost of being tiring to the ears. Sinc-M and M(x) are a little smoother, which can sometimes be a good thing and sometimes I prefer Sinc-L, depending on the recording.

Have you tried poly-sinc-guass-long + poly-sinc-gauss-hires-lp (for oversampling 1x and Nx, respectively)? This one sounds very different to me than the Sinc filters, I used it for a good while before recently switching back to Sinc but I’m already missing what this poly gauss filter offers. To my ears it makes sounds flow more freely within the soundstage, making the Sinc filters sound a little “forced”/congested in comparison. On certain recordings the difference can actually be stunning. The only thing I sometimes miss is the tactility/vividness of the Sinc filters. When it comes to focusing on particular sounds, e.g. soaking up the sound of the electric guitar, the Sinc ones seem to convey more tactility and body behind those sounds which can be very satisfying and more life-like. But then the poly-gauss filter is much better at conveying the entirety of the musical experience, in terms of making me believe that the sounds are individually located on an imaginary stage, as opposed to coming at me in a more aggressive and often unnatural way.

The poly-sinc filter with the tactility/body/vividness of the Sinc-L filter would probably be my ideal filter but I guess it’s one of those issues where trade-offs need to be made and we can only attempt to have the best of both worlds. I’d be keen to find out what your perception of this filter is.

[KML]

I have tried the gauss-long, but not extensively. I will give it another try soon, thanks for the recommendation, @MistaLovaLova.

For now, my new point of reference is IIR (with Gauss1 dither). Have you tried that one? It is a filter that I overlooked in the past, mainly because it sounds best at x2 rates, and HQP doesn’t allow to personalize the input beyond 1x/Nx like Roon does. (I wish it could!).

With IIR, anything “digital” gets out of the way. To my ears, even sinc-M sounds a little bit thin or metallic by comparison! I’d be curious to know if you (and others) hear it that way too.

There is also the minphaseFIR filter which seems like a good compromise between “analog” and “digital hi-fi”. It sounds pure and undistorted for both timber and harmonics, and offers a relaxed but engaging listening. By “digital hi-fi” I mean a sound that is detailed, sharp, holographic (like long and XL versions of gauss, for example), but still sounds a bit thin or digital.

Lately, I have been very sensitive to unnatural echoes/reverb created by linear or long filters. I now prefer to trade off a little bit of detail (after all, those chip-based dac have plenty of that) for something more real and analog. Although I mainly listen to classical and jazz music, I now tend to prefer minimal phase filters. Poly-sinc-long-mp (or even Gauss-Xla and ETX3), for example, have more “echo” than IIR which makes them sound a bit more “3d”. At first, that’s very engaging. But they tend to present the contour of notes and instruments, rather than the inside. IIR sounds flatter, but also more grounded, full, organic, less “metallic” and fatiguing. Anyways that’s how I hear things now, and I might very well review my judgment in a week or two!

p.s. The problem with sinc-L, is that it is non-apodizing. After doing the gauss-XL/XLa comparison test, I can now understand why I need apo.

[dabassgoesboomboom]

@KML

p.s. The problem with sinc-S, is that it is non-apodizing. After doing the gauss-XL/XLa comparison test, I can now understand why I need apo.

Per the manual, sinc-S is apodizing

[jussi_laako (Jussi Laako (Signalyst))]

[quote=“KML, post:72, topic:244185”]In other words, length or speed (of a given filter in time-domain), and steepness (of the cutoff frequency) are always related. I am putting it together correctly?
[/quote]

Yes, that’s correct. Except speed is wrong term there, “speed” (as in speed of signal change) in time domain is function of bandwidth in frequency domain.

@KML

I am trying to make sense of what I hear. Suppose I use a slow/super slow filter : I clearly perceive that the trebles are somewhat attenuated. It often gives me the impression that the sound is a bit less transparent, somewhat less “sharp” or strident in the highs.

Yes, when you go too slow roll-off, that happens. And when you go over the board to too long, you begin to have artificial “sheen” around the transients which is due to the prolonged time domain response (ringing).

@KML

I guess my main question is : how are the attenuation of the passband and that of the stopband related? Does a super slow filter necessarily have a lower cutoff frequency (thus rolling off some of the trebles in the passband?)

No, slow roll-off filter can be flat to Nyquist frequency. But it also means it leaks a lot of image frequencies that generate intermodulation and “harshness” or “edginess” to the sound. It begins to sound rough and coarse. Or some could call it “digital”.

@KML

From what I see in a typical representation of digital filters, a steep slope is most of the time correlated to a higher frequency of the passband and to a higher cutoff frequency. Is that correct?

That is design decision. Cut-off frequency is independent of the other parameters.

@KML

Conversely, I see slow filters start to roll-off trebles in the passband. In plain words, when I hear more trebles and transparency, is it mainly because of more stopband attenuation, or because stopband attenuation inevitably means more passband attenuation? Am I completely off base?

No, passband roll-off and stopband attenuation are unrelated. But typically on a DAC chip filter where there are limited computing resources, a lot of trade-offs are made because number of taps and precision is very limited. So stopband attenuation is traded for increased steepness and vice versa. While HQPlayer doesn’t consider how much something takes computing resources, it is just purely designed for certain design objective, and then you need to have fast enough computer to process it (in any case several orders of magnitude more than what on-chip DSP can do).

What is relevant is the overall filter function, which is much more complex than what can be described by couple of simple parameters.

[KML]

Yes, thanks. That was a typo, I meant sinc-L. I didn’t try sinc-S extensively, I but I know sinc-M/L very well. (I just corrected the post).

[MistaLovaLova]

@jussi_laako

Yes, when you go too slow roll-off, that happens. And when you go over the board to too long, you begin to have artificial “sheen” around the transients which is due to the prolonged time domain response (ringing).

That would explain why the Sinc-L filter can feel unnaturally pronounced, i.e. its punchiness and vividness (whilst impressive for certain genres) generally coming across as artificial. And why the M(x) is a little bit less so (less steep cut-off) whilst still being very much in-your-face compared to the gauss-long.

Am I right in thinking that my go-to poly-sync-gauss-long is a trade-off between those two approaches? I can see in the manual that it is extremely high attenuation, I’m wondering if it’s still a sharp/very sharp roll-off/cut-off filter or if it would be classed a slow one. And whether the difference between gauss and gauss-long would be in their steepness.

[jussi_laako (Jussi Laako (Signalyst))]

@MistaLovaLova

Am I right in thinking that my go-to poly-sync-gauss-long is a trade-off between those two approaches? I can see in the manual that it is extremely high attenuation

It has more attenuation than sinc-M.

@MistaLovaLova

is a trade-off between those two approaches?

It is a different type of filter, not really a trade-off between sinc-L and sinc-M which are both extremely long.

@MistaLovaLova

I’m wondering if it’s still a sharp/very sharp roll-off/cut-off filter or if it would be classed a slow one.

gauss-long is on sharper side of the middle, but something I consider very much sensible somewhere in the middle between sharp and slow. If you take 44.1k RedBook content, it is flat to 20 kHz, but reaches full attenuation by Nyquist (22.05 kHz).

@MistaLovaLova

And whether the difference between gauss and gauss-long would be in their steepness.

gauss is intended to be precisely in the middle between short/long. Primary difference is in steepness (length), but there are some differences in other parameters as well so that it fulfils it’s design criteria. For each filter, I have certain set of goals they are designed to meet.

[MistaLovaLova]

Thanks. And apologies, I didn’t make myself clear, when I was referring to a trade-off between the two approaches, I wasn’t referring to the Sinc-L and Sinc-M(x) filters; what I had in mind was the very sharp and very slow roll-offs/cut-offs. I should have made myself more clear.

It’s nice to learn a bit more about the technical considerations behind creating these filters. This seems to reinforce what my ears have been hearing for a while now - that the poly-sinc-gauss-long does seem to be the current Goldilocks digital filter, best all-rounder.