Sound Quality Improved PDF Print E-mail
Thursday, 03 August 2017 15:47

432 EVO is continuously searching for ways to improve the sound of it's award winning music servers.

We have looked deeper into our own DSP recipes, and found several ways to drastically improve the sound quality of the 432 EVO music server. We have decided to publish these improvements as a free software update.

The basics

Let's start with some 432 EVO basics. All 432 EVO music servers render any PCM file to 32 bit into the desired output resolution set by the 432 Hz plugin:

This is set using the Configure Player form:

When a 432 EVO music server leaves the factory, it is set to 32/192, which is a resolution most DAC's currently can handle. It is also the maximum resolution of the internal DAC in the 432 EVO ESSENCE music server. Looking at the Mytek Brooklyn, the display confirms that it actually receives 32 bit and 192 Khz:

For backwards compatibility with very old 16/44.1 and 16/48 DAC's, when you select 44.1 or 48 Khz, the output is set to 16 bit, otherwise 32 bit.

This article will describe why 32 bit processing does not make much sense.

The DAC's own noise floor


Until the end of 2017, all EVO's would pass the 32 bit output from the plugin via USB to the DAC (or in case of the ESSENCE, to it's builtin DAC). The issue here is that no 32 bit DAC actually exists.

A perfect 24-bit DAC would have an SNR of close to 144 dB, but there is no delta-sigma DAC out there that comes close to that. There are DACS specified as 32-bit, but their SNR is no better than the corresponding 24-bit version, which in turn can't do anything close to 144 dB.

Let's look at some measurements done by Archimago of typical 24 bit DAC's:

As you can see, most DAC's don't reach the actual noise floor of around -144dB. 

So while the DAC receives 32 bit in the S32_LE format, it first throws  away the 8 LSB bits so the data passed to the actual DAC chips is 24 bit. From this 24 bit, most DAC's reach the analog equivalent of 20 ~ 21 bits, as only a few DAC's exist with a noise floor that corresponds to true 24 bit. Some super expensive DAC's claim to be around 26 bits, but those measurements are still debated. For our own testing, we use Metrum's flagship Adagio DAC which claims to be a true 24 bit DAC. The good news about this DAC is that it does not filter in the digital domain, so any audible DSP difference as presented here, is enlarged on this DAC.

First solution: Less is more!

So the first trick was to render into 24 bits resolution instead of 32 bit. Even if this is being passed as S32_LE data to modern DAC's, when the DAC throws away the lowest 8 bits, those bits were already blanked anyway. Smart DAC's such as the Mytek detect there's only 24 bits of data inside a the 32 bit wordlength send over USB. One vendor actually confirmed to us his excellent DAC reaching a noise floor of 145dB only uses 24 bits from the 32 bits presented over USB and described to me how they throw away 8 bits.

So the first version of SQi was actually called Deblur:

The first variant (shown in blue) forces the output to render in 24 bits instead of 32 bit. The second variant uses the same bit depth, but with a less steep filter. These 2 options added a sense of more resolution, appreciated by most customers. This was tested in february 2017.

Second solution: 24 bit with tweaked filters including linear phase, minimum phase and intermediate phase!

In Q4 2017, deblur was renamed to SQi, "Sound Quality improved", and more recipes were added. 

When you set the BASE and TARGET frequencies to 440, it acts as a high-end upsampler with much higher precision than what is available in most DAC's. When you set the TARGET to 432 or any other frequency such as 430.5, it will change the frequency of the A, based on speed shifting instead of pitch shifting, as pitch shifting sounds fake. Except for a very subtle tempo change, 432 EVO's 432 Hz method has no audible artefacts. This is all explained in our FAQ.

We now also offer the new A=430.5 target frequency endorsed by another company who also believes in C=256 Hz sounding better, but using a different tuning system so A becomes 430.5 Hz (scientific tuning) instead of 432 Hz (Pythagorean tuning). 432 EVO offers both tuning, and several variants such as A=444 (Horowitz tuning) and A=426.6 (Diatonic Scale Tuning).

The processing power of the 432 EVO allows any DSD input file up to DSD256, any PCM input file including 16/44.1 up to DXD and 24/384, to be output in any PCM output resolution your DAC supports, in any support BASE,TARGET frequency pair, or with just upsampling (where both are set to 440), while keeping power consumption low and keeping the machine cool.

Linear phase vs minimum phase

To understand the SQi modes, we need to explain minimum phase vs linear phase first.

Our hearing is very senstive to sounds before the actual transients. The default upsampler before SQi was offered uses linear phase. When a pulse is going through such upsampler, there will be ringing before and after the peak. This ringing will be symmetric, as all frequencies pass through the filter with the same delay. The fact that there's a signal before the actual transient results in subtle loss of dynamic range.

When using minimum phase, there will be no pre-ringing, and all ringing will be after the actual pulse. This post-ringing will be buried in the decay of instruments and reverb of the room. Our hearing is very sensitive to pre-ringing, but not post-ringing.

A special case: minimum phase with one cycle post-ringing

The above impulse response plots have long pre- and/or post-ringing tails. This is necessary, as the steeper the filter, the longer the ringing. A special variant exists, where one allows aliasing in the upsampler to create a very shallow filter. This filter was already described in 2009 by Ayre. While 432 EVO does not use the exact filter Ayre has implemented, it has a very similar transient response as what both Ayre and MQA are implementing:

Image (c) Ayre, screenshot from PDF used under fair use.

Looking at the plot of standard minimum phase, vs the one cycle variant, we get a very similar response as both Ayre and MQA:

While these filters are based on the apodizing filter of Craven, Ayre further modified these filters to feature one cycle of post-ringing, which they call the best of both worlds. No IP was used from Ayre products. These filters were found independently by manipulating passband and stopband parameters in existing resampling libraries, and several similar variants exist.

As you can see, when upsampling a 96 Khz impulse file to 192 Khz, the transient takes an estimated 17 samples. At 192000 samples per second,  this leads to 88,5 microseconds.

Now looking at the Sound on sound article for MQA's claimed time domain accurary, we can observe a very similar impulse response (blue plot):

Image (c) Sound on sound, used under fair use.

All the modes explained

SQi Disabled

This is the default mode since 2013, where we render in 32 bit and leave the 32 bits as-is to the DAC, which typically throws away 8 bits.

When using very high, this is based on the "v" recipe of the SoX resampler.
When using ultra high, this is based on sox's libsamplerate emulation for libsamplerate's "best sinc", which is very efficient code.

We prefer Ultra High instead of Very High as the SoX "v" recipe can sometimes sound a little harsh on very revealing gear.

SQi Linear phase normal

This uses the same resampler as SQi disabled, but with 24 bit output instead of 32 bit. So this uses the first pillar on which SQi is based: less is more.

SQi Linear phase with slow rolloff

Similar to linear phase normal, but with a less steep filter. Our preferred linear phase filter.

SQi Linear phase high precision

Similar to linear phase normal, with steep filter, and forced internal 28 bit precision. An alternative to linear phase normal.

SQi Minimum phase

Our standard minimum phase filter, without aliasing and no pre-ringing.

SQi Minimum phase high precision

Our standard minimum phase filter, without aliasing and no pre-ringing and forced internal 28 bit precision.

SQi Time domain

Minimum phase filter, with excellent time domain impulse response, but at the cost of aliasing and ringing. This will make certain instruments more tight, but may lose some depth and decay, as the post-ringing is shortened. Can be fun with electronic music and EDM. Makes the bass kick more.

SQi Time domain high precision

Same as Time Domain but with forced 28 bit internal precision.

SQi Archimago Intermediate Phase

This filter is none of the above, and tries to be the best of everything.  This filter has partial pre-ringing. Full technical discussion here.

This filter has 28 bits of internal precision, and always forces the "v" = "Very high" recipe of the SoX resampler.

All the Archimago based filters ignore the setting for recipe, and override them manually:

SQi Archimago imp + evo stage 1

This is 432 EVO's version of the filter, tweaked to have more soundstage. Based on sox "v" = "Very high" recipe.

SQi Archimago imp + evo stage 2

This is 432 EVO's version of the filter, tweaked to have more soundstage. Based on libsamplerate emulation (aka recipe = Ultra High).

And what about internal precision?

To use SQi , we do not recommend to use internal precision set to 2x or higher. Set it to 1x!

So suppose you have a 24/96 dac and you would set DAC output resolution to 32/96 and internal precision to 4x, it would tell the player to upsample to 32/384 (combined with 432 Hz processing if “From:” was set to 440 and “To:” set to 432), and at the end of the processing chain, it would downsample it back to 32/96. This way the 432 Hz plugin would have more precision, but this theory is no longer supported for 24/192 and higher DAC’s.

Internal precision actually renders to a virtual soundcard with 2x, 4x or 8x the resolution of the actual wanted DAC resolution, and at the end it downsamples it back to the set DAC resolution using a very high quality sox recipe. This was invented to have more precision for older DAC’s limited to 16/48, 24/96 … and is no longer recommended when using SQi.

So the optimal value when using Sqi is 1x for internal precision:

When changing the filters, just press the submit button on the Configure Player form, and the player will restart the current song. When using Roon, the zone will reappear immediately, but you may press the play button in the Roon app to continue playing. When reloading the player config, Roon will pauze the player.

In Munich we demonstrated the 432 EVO in 440 mode, with SQi active:

432 Hz processing and MQA?

MQA is not compatible with any vendor selling DSP solutions including room correction, crossfade, digital equi, bass management. With MQA you get the right to listen to an approximation of the studiomaster, but you'll never get the actual PCM data that the master engineer was working on. As the process is lossy, the original PCM that was used in the studio is never transferred via an MQA distribution file. The Chaos Computer Club has a good presentation on these issues. MQA will never allow us access to the first + second unfold to run our 432 Hz processing on top of that. Brian Lucey is a Grammy award winning engineer, who has also found MQA to sound different from his own masterings. Furthermore he found his masterings to be encoded into MQA without permission. So MQA's policies are questionable to say the least.

But there is a way around.

First the sound of the true master file which was used to encode the MQA files can be recovered using the sox minimum phase recipes, which we offer as a free DSP upgrade (part of SQi). We did a lot of testing with befriended hifi professionals and send them our upsampled MQA files using sox minimum phase, and the originals from which they were encoded. They could not hear the difference:

Only our most expensive system could reveal very small differences, which will not trigger with most of our customers.

Furthermore, the time domain filter we present reaches a very similar impulse response as MQA's renderer, which is basically an upsampler / ditherer. Remember that MQA's first unfold recovers some ultrasonics up to 48 Khz by embedding these as a lossy version in the 8 lowest bits of a 24/44.1 or 24/48 distribution file. The second unfold does not create new data, but upsamples the first unfold while allowing aliasing.

Here is's 2L-053 demo track in the original studiomaster (blue plot) vs the first unfold (red plot) and second unfold (green plot):

The first unfold actually is true to the original spectrum but in a band limited fashion, as it stops at 44.1 Khz nyquist. files are all multiples of 44.1 Khz and edited as DXD.
The second unfold which upsamples + dithers the first unfold just creates a totally new response not in the original.

Furthermore in a recent stereophile article, MQA did not perform much better than redbook in terms of noise floor, and the reviewer could not hear the higher noise floor compared to the DXD studiomaster. The MQA version had 20dB more noise than the original, but this was mainly in the ultrasonics range, see fig2 in the link.

Most likely because redbook's noise floor is more than enough, and extra garbage in the ultrasonics range is not a problem for the human hearing. We recorded concert halls with an RT60 of 2,03 secs. RT60 is the time for a crescendo to attenuate back to background noise levels. 60dB is 10 bits worth of headroom. Add an extra 10dB for some headroom below the noise floor, and we can get away with 12 bits of resolution. MQA typically reaches 15 to 17 bits of resolution, but still gets away with burying 8 bits of lossy HF spectrum encrypted and DRM'ed in the LSB parts of the MQA distribution files.

Here's K's Choice in choir, as recorded in 16/48. From an a capella song with just the 3 band members singing in this 940 seat hall, to the full 240 piece choir singing at maximum volume, it all fits in the headroom of 16/48. Analysis of the wave files and playback on a very high end set confirmed this very special concert hall (deSingel in Antwerp) has optimal acoustics and our good old redbook format is sufficient. We later recorded another event in deSingel with professional gear from our good friend and cinematographer Kommer Kleijn to confirm these findings. Kommer is a researcher into psycho-acoustics and also teaches sound reproduction at the Brussels film school. He is also an inventor.

432 Hz is still a winner

In all our testing, no decoding + the sox upsampler + 432 Hz processing still wins from full decoding in an MQA dac and leaving the music in A=440.

Remember that 90% prefers A=432 Hz over the original in A=440 Hz. The difference between highres and MQA is much less drastic. In a blind test done by Archimago, the effect is much smaller than the 432 Hz effect. The effects were like 45% highres vs 55% mqa. 

Bitperfect output? MQA passthrough?

Some customers prefer bitperfect playback. 432 EVO can be used as a bitperfect source. In case of the all new 432 EVO MASTER, the performance is so high it can compete with any other digital source no matter the price point, and this source won best of show awards in Munich 2017.

To use the 432 EVO as a bitperfect source, just set the DAC Output resolution to "Disable Plugin". This will start a bitperfect player.
This will also pass MQA files to an external DAC which can do the full decode. Don't forget to set the volume to 100%, as volume control implies the output is no longer bitperfect.

432 EVO gives you a lot of choices and does not enforce one way of playback. You can play with many parameters on the Configure Player page to tune the sound to your own taste.

SQi Release  = 14 jan 2018

We have released SQi with 10 digital filters instead of the default 2 on 14 jan 2018. Happy updating.

Last Updated on Sunday, 14 January 2018 21:54