AuraEngine: Perfectionist off-line mega-tap digital filtering - with hybrid-phase transient detection & blending.

As you know, while digital oversampling filters have been around for ages (4x oversampling available back in the '80s), audiophiles probably got more interested in digital filters thanks to Bob Stuart, and Meridian by 2009 talking about "apodizing" filters. DACs started allowing different filter options. And spin-offs like MQA took things too far utilizing poor designs through the 2010's (and still looking for a reason to exist).

Years ago now, I spent quite a bit of time talking about digital filters across articles like this and contemplating things like pre-ringing here. There were also discussions of using the intermediate phase setting as a compromise. And we even ran a blind test to compare strong pre-ringing vs. no-pre-ringing minimum phase settings.

While I think I've exhausted my time exploring such things at the base level, ever-more-precise filtering options have been available for audiophiles to use. HQPlayer has many realtime software-based playback settings including DSD transcoding. Companies like Chord/Rob Watts have taken their stand on using ultra-tap-length linear phase filters (to the point of extreme IMO).

Off-line upsampling of high precision can also be achieved using software like Pan Galactic Gargle Blaster (PGGB) in 2021 -  I see they have a PGGB+ version now with DSD and up to "ludicrous" 256-bit processing.

Well friends, the other day I was contacted by Toxa Dev about his new ultra-accurate upsampling system called AuraEngine (current version 1.2.2) available on Github; full source code published and freely released for non-commercial use. You can also download and run the pre-built Windows portable app here. Notice that there are three downloads available: Starter 1M-taps, Standard 10M-taps, and Reference 30M-taps. Pick your level of obsession err... weapon against weak upsampling!

In a nutshell, this app will take your stereo 44.1/48kHz PCM file and upsample it up to 705.6/768kHz (2x all the way to 16x available), floating-point 64-bit "double precision" (15-17 decimal digits), with no clipping (intersample overs detected with -0.5dBTP ceiling), up to 30M-taps filtering. It can achieve better than -220dB stopband attenuation, and transition width can be down to 0.05Hz!

[Mathematically beyond Chord's 56-bits, 1,015,808 taps of the Hugo M Scaler, eh? And the Chord Quartet's 4M taps suddenly isn't a big number, but an asking price of £25,000 is still fairly large, it's realtime playback so there's that. 😉]

Furthermore, the app can take advantage of your GPU's computational power (using Vulkan API). And even better, if you want to spend more computational energy, it can apply a "Hybrid-Phase engine". Cool!

Let's talk about that hybrid engine. Normally, a linear phase low-pass filter is best for high-fidelity playback because it does not impart a phase shift / excess group delay at higher frequencies:

Unfortunately, at times on some recordings with sudden transients (typically with dynamically compressed material, poorly band-limited, and more likely than not synthetic content as discussed), we can see pre-ringing showing up. AuraEngine has an ability to scan the signal, look for portions where transients could cause pre-ringing and seamlessly transition to minimum phase filtering at those strong transient points. Thus, achieving a "best of both worlds" result - no phase shift for the vast majority of the playback, and no pre-ringing during brief transients.

Let's show this in action using a synthetic "obstacle course" to see how and where it works:

Notice what I'm doing here. I start off from a very brief period of silence right into a square wave. Then transition to a sine that fades, then a sudden loud square waves (transient) at ~0.5s, and then a continued run of square waves (clearly non-musical material!).

The app has a transient detector with an adaptive gate. Due to the very short ~20ms silence before the first square waves hit on the left, it doesn't trigger the switch. However, once the software sees and follows along with the short run of sine waves that then fades, when the loud square wave enters, we see that this triggers the transient detector and pre-ringing has been suppressed for a short period of time. Notice that the subsequent square waves to the right goes back to linear phase with pre-ringing since the actual transient has passed.

For the more technically curious, Toxa Dev kindly sent me this inside look into the switch threshold and envelope of the hybrid-phase transient mechanism with values like the 30ms switch-opening prior to the impulse:

The take-home message is that this hybrid-phase mechanism is not a system meant to prevent all pre-ringing. It's a strategic mechanism to detect and mitigate pre-ringing in actual music where every once awhile a strong transient could cause brief pre-ringing, potentially "blurring" the impact of that transient. Otherwise, music is played back with a clean, linear phase filter without time-domain distortion.

Utilizing the "reference" 30M-taps, we will clearly see very strong "brick wall" filtering (here's Jennifer Warnes' "Bird On A Wire" run through FFT, dithered 24-bit output):

In this example, 20kHz "apodizing" applied; starts attenuating below Nyquist. Can be turned off. There's also an adaptive apodizing feature that can detect poor recordings with source pre-ringing (read more here).

And yes, if we feed a severely dynamically compressed signal into the upsampler, it will handle bad intersample overs - here's Bastille's "Pompeii (Audien Remix)", an atrocious DR3 track:

Notice the strong intersample overs; some going above +3.5dBTP in the original file which would likely cause clipping on many DACs. AuraEngine ensures that the intersample peaks do not go above -0.5dBFS.

Of historical interest, we have seen a hybrid phase system before - decades ago! Back in 1993, the old $15k Meitner IDAT DAC also incorporated a transient detection system and would flip between a minimum phase IIR filter and linear phase FIR filter employing four Motorola DSP56001 chips. Pretty cool stuff for its time. I do wonder though how accurately that detection system performed to signal the filter change and the precision of that switch (ie. what anomalies seeped through at the margins?). Fascinating bit of history and I remember being quite curious about the Meitner device when I first read about it years ago. In contrast to that old realtime system, AuraEngine performs parallel linear and minimum phase conversions, splices the best filter output, optimally aligned at 0-crossover with some micro-fading. This dual conversion is why using hybrid-phase blending doubles the processing needs.

So, you must be wondering - "How does this sound, Arch?"

Well, I think this is the best I've ever heard the DR3 "Pompeii (Audien Remix)", sound. 😁 Clearly all the intersample overs have been addressed, and the transient detector fired more than a few times throughout the song to suppress pre-ringing. The most important thing is to not hear any unusual distortion in the final output - I did not hear any issues.

A peek at the transient strength across time for "Pompeii" recorded
by AuraEngine in the metadata produced during processing.

As for much better recorded music like say Jennifer Warnes' Famous Blue Raincoat (20th Anniversary), DR12, what can I say, it sounds as good as I would imagine precise PCM upsampling could achieve. Clear vocals, no nasty sibilance, precise bass and guitar plucks, lively fiddle, lovely soundstage, subtle effects like the crickets, quiet background - as it all should be.

Notice how different the transient strength graph looks for
Jennifer Warnes' "Ballad of the Runaway Horse". On average
~50% the number of onset events/minute compared to "Pompeii".

I told Toxa Dev that while I can't put a finger on any specific subjective sonic change, the precision his app performs at is clearly real.

Bottom line...

For you digital filter perfectionists, give this upsampler a go! Instead of just brute force upsampling using static parameters, it's got both brawn and brains to know when to intervene and when not to. Phenomenal work, Toxa Dev.

Feed your favorite 16/44.1 CD track from the very best mastering you have (poor masterings might even improve more!) into this program. Upsample to at least 24/352.8 with the 30M-tap setting (0.05Hz transition width), turn on the hybrid-phase blending. Since most DACs will have turned off their internal oversampling already by 352.8kHz, play that upsampled file to a compatible hi-res DAC and you'll be treated to the sound of a contender for the best PCM upsampling you'll likely ever hear - and it's free!

Of course, make sure to invite your friends with Golden Ears over for a listen! Rob Watts should be more than a little impressed. 😲

[For more technical information about AuraEngine, make sure to check out the "15-measurements.md" document. Public discussions on Github and private E-mails can be sent here: auraengine.dev@gmail.com.]

--------------------

Okay dear audiophiles, some fun music to take us out for the last few days of "Vacation" - Dirty Heads (from Swim Team):

With that, I wish you a great rest-of-August. With September around the corner, and kids are getting ready for school which means it'll be back to busy-as-usual around here.