AuraEngine Converter

[Erogen]

Audacity > Analize > Plot Spectrum

 

[formdissolve]

Hmm, on Linux it doesn't look the same for me. I can try it on Windows later. I've had success on a few albums so far with de-clipping, but I haven't yet critically listened to see if I can hear a difference.

I did notice one album so far that has a much busier spectrum - maybe it's the hybrid filter or something. I forgot to see if there was clipping in the first place, but here's a track with before and after:

 

[Erogen]

I use the Foobar2000 program; everything displays nicely.

 

[ToxaDev]

1.3.1 is out, and it is housekeeping rather than anything you will hear. Two
of the heaviest stages — the declipper and the crosstalk canceller — were each
running on a single thread. They aren't now. No arithmetic was reordered, so
the files come out as they did before; they just take less of your evening.
The release notes say which parts had to stay sequential and why, for anyone
who cares about that sort of thing.

There is also a browser version now. Same engine, same pre-computed filters,
up to thirty million taps — nothing to install and nothing to pay. It exists
for the people in this thread who aren't on Windows, and for anyone who would
rather hear what it does to one track before putting a gigabyte of filter
coefficients on their disk. Eight tracks at a time, up to eight minutes each;
whole albums are still the desktop build's job.



Windows builds, filters included, four sizes from 7 MB to 967 MB:
github.com/ToxaDev/aura-engine/releases/tag/v1.3.1

 

[Erogen]

AE 1.29
AE · 44.1k→176.4k · Kaiser 30M · f64 · SUB20 · AA · HP

Time: 3179s (14 track)
Intel Ultra 7+ GPU RTX 5060TI (C: 3008MHz/ M: 1788MHz)
Memory 32GB (available 25 GB)

1 program instance processing speed: 28.13x + 6.02%
2 program instances processing speed: 31.47x - 9,18%
Two instances of the program mean 2 × 4 active threads. The program reports waiting for available memory, causing momentary pauses. It therefore seems that a single instance with 5 or 6 threads would be optimal.
However, I believe that two instances with (2 ×) 3 active threads will yield the best result. – 34.55x

Ver. 1.31

Time: 3179s (14 track)
Intel Ultra 7+ GPU RTX 5060TI (C: 3008MHz/ M: 1788MHz)
Memory 32GB (available 25 GB)

[AE · 44.1k→176.4k · Kaiser 30M · f64 · DC·ISP·SUB20·AA·PFR·HP]

1 program instance processing speed: 26.06x
2 program instances processing speed: 33.11x

 

[Erogen]

Before

After

[AE · 44.1k→176.4k · Kaiser 30M · f64 · DC·ISP·SUB20·AA·PFR·HP]

 

[Erogen]

Major changes—it plays differently.

 

[RandomEar]

See what simply lowering the volume does for you:

That's what this program does:

I just stumbled upon this thread. I think the explanations by @ToxaDev have been very good! But just in case anyone wondered why these two plots still look different:

The first plots shows the manually attenuated recording to avoid ISOs - let's assume by -3 dB (the exact value isn't important to follow the argument). It still looks like a lawnmower ran over it. The second plot shows the recording after using AuraEngine which mainly did two things to it: Attenuate based on true peaks and apply a reconstruction filter which would otherwise run in the DAC. So this comparison shown here is misleading, because it does not compare what leaves the DAC. If you compared both versions at the DAC output, they would look much more similar because the DAC's reconstruction filter would als add values above -3 dB to the first version of the file, exactly at those positions where ISOs occurred in the unattenuated original. So at the DAC output, the first version would also lose that "lawnmower look". There would still be differences between the manually attenuated version and the AuraEngine one due to the different reconstruction filters, slight differences in attenuation and - of course - bigger differences if you used any of its DSP functions. But the qualitative difference would not be as huge as the cited plots above suggest.

Also, since the question of reconstruction filters came up: I would say that if you correctly upsample (>300 kHz) and band-limit the file before sending it to your DAC, this would be pretty much the only case where NOS mode in the DAC could be considered a valid option. Due to the very high sampling rate, none of the slow roll-off filters would cause an audible reduction of treble either, so I would say that the filter choice in the DAC would be pretty much irrelevant here.

 

[ToxaDev]

@Erogen You're right, it plays differently, and that's my fault: I never posted here about 1.3.0. That's the release where the declipper from my first post went into the app, switched on by default. 1.3.1 changed nothing on top of it, but coming from 1.2.9 you get both at once.

Your table shows what it does. On the disc the peaks are cut flat, and the declipper draws them back. The redrawn peaks are taller than anything on the disc (your loudest moment moved from 1:35 to 0:23), and the file still has to fit under the ceiling, so the whole track comes down, here by 7 dB. That's why DR went from 7 to 12: the peaks now stand further above the average. LRA stayed put, 4.66 → 4.65, so the loud and quiet parts of the song are as far apart as they were. A lot of the "different" is likely just that it's quieter, so compare them with ReplayGain on. If you want 1.2.9's behaviour back, turn off DC and ISP.

The 18 kHz wall isn't new. It's the adaptive apodizer (AA), it was in your 1.2.9 files too, and you can switch it off.

The one thing I can't match is the bottom end. With SUB20 everything below 10 Hz should be gone, and in a test here with your exact settings it is. Your After still shows the disc's own rumble there, and it sits at the same height as Before, although your table says the file is 7 dB quieter. Is that plot straight from the AE file, or did it go through something after? If it's straight from AE, could you post the "Subsonic filter SUB20" line from the log?

 

[Erogen]

"Subsonic filter SUB20" line from the log?

[10:09:27.862] [CONV] Subsonic filter: 20 Hz
[10:09:28.107] [CONV] Subsonic filter SUB20: 29813 taps (0.68 s), linear phase, flat from 20 Hz, ≥100 dB below 10 Hz; removed RMS L -61.1 dB / R -62.4 dB re full scale

 

[ToxaDev]

@Erogen Thanks for the log line. It says SUB20 ran on that track, so the file AE wrote shouldn't have anything left under 10 Hz. Here's what that looks like on a clipped CD of mine through your settings, plotted with the same math Audacity uses:

Top is the disc. In the middle is the 1.3.1 file with SUB20: below 20 Hz the plot drops off a cliff. At the bottom is the same conversion with SUB off: the shelf runs all the way to the left edge, like on the disc.

Your After looks like the bottom one. It also sits at the same height as your Before, while a converted file comes out lower: 3.7 dB here, 7 dB on your track by your own table. So I think that plot is of a different file, maybe an older conversion. Worth knowing: Audacity adds up every selected track of the same sample rate before it plots, so if two conversions were open and selected, you get the sum of both. Could you open just the file with SUB20 in its name, alone in the project, and plot it again at 65536? If the shelf is still there, I've missed something and I want to know.

 

[Erogen]

The De-Sub filter works. I must have mixed something up.

 

[Erogen]

I’m not an expert on this. I wanted to ask why the frequency range is cut off here at around 18 kHz—aside from a few isolated peaks.

 

[Erogen]

Adele - 19 (2619s)
Intel Ultra 7 +GPU RTX 5060TI
AE Ver.1.31 [AE · 44.1k→176.4k · Kaiser 30M · f64 · DC·ISP·SUB20·PFR·HP]
1 x AU - 31,18x
2 x AU - 35,39x (1-6, 7-12)

The songs are very good, but recording and final mastering aren't the best. However, it didn't matter—the album was a huge success.

 

[ToxaDev]

I’m not an expert on this. I wanted to ask why the frequency range is cut off here at around 18 kHz—aside from a few isolated peaks.


View attachment 560104

@Erogen There are two things in your picture, and the second one was a real bug. Thanks for spotting it.

The flat ceiling at 18 kHz is the adaptive apodizer (AA). A CD's audio goes through a very steep filter just below 22 kHz, and a filter that steep rings: a faint tone up there that starts a few milliseconds before every sharp hit. AA looks for that ringing, and when it finds it, it trims the top with its own filter, one that doesn't ring before the hit. The more ringing it finds, the lower it cuts, and 18 kHz is as low as it goes. What it takes is whatever the disc had between 18 and 20 kHz, about 50 dB below the bass on your plot. If you'd rather keep that, switch AA off.

The thin spikes above it came from Hybrid-Phase (HP). HP renders the track twice, through two versions of the filter, and switches between them around the hits. That switch was far too abrupt, 0.08 ms, and every one left a tiny tick with energy at every frequency, so it poked through above the cut. On a clipped CD of mine with your settings, the strongest reached about −43 dBFS. They're ultrasonic and I don't claim anyone heard them, but they shouldn't have been there.

1.3.2 fixes it. The switch now takes 1 ms on a smooth curve, at the same points as before; nothing else in the conversion changes. Same CD, same settings, before and after:

Above 24 kHz it went from −42.9 dBFS to −103.0. The cost is that for that millisecond both versions play together, which takes about 0.3 dB off 14–21 kHz right around each switch; below 8 kHz nothing changes. Only files made with HP on are affected; a fresh install starts on TFS, which never had it. The web version has the same fix.

github.com/ToxaDev/aura-engine/releases/tag/v1.3.2

 

[MaxwellsEq]

The flat ceiling at 18 kHz is the adaptive apodizer (AA). A CD's audio goes through a very steep filter just below 22 kHz, and a filter that steep rings: a faint tone up there that starts a few milliseconds before every sharp hit. AA looks for that ringing, and when it finds it, it trims the top with its own filter, one that doesn't ring before the hit. The more ringing it finds, the lower it cuts, and 18 kHz is as low as it goes

Would it have gone below 18kHz if it was unconstrained? To a certain extent, you are reproducing the "roll-off" present below 22kHz with slow filters. One criticism of such slow filters (aside from poor attenuation above fs/2) is that they artificially reduce HF that's actually on the disc.

As I understand it, and am happy to be corrected by experts, the (pre) ringing you refer to has a fundamental frequency of fs/2, so for CD, the ringing has a frequency of 22kHz and is thus inaudible to just about everyone over 12 years old.

 

[ToxaDev]

Would it have gone below 18kHz if it was unconstrained? To a certain extent, you are reproducing the "roll-off" present below 22kHz with slow filters. One criticism of such slow filters (aside from poor attenuation above fs/2) is that they artificially reduce HF that's actually on the disc.

As I understand it, and am happy to be corrected by experts, the (pre) ringing you refer to has a fundamental frequency of fs/2, so for CD, the ringing has a frequency of 22kHz and is thus inaudible to just about everyone over 12 years old.

Only a little, and only if it had measured the ringing below about 19.8 kHz; then it could have gone down to about 17.3. With ringing as strong as on that album, the preset lands on 18 kHz anyway. The floor is there because an early version followed a misread down to 17.6 kHz.

On the other two points you're right. The ringing sits at the edge of the disc's filter, around 20–22 kHz on a CD, and I'm not claiming anyone hears it. And yes, it's the slow-filter trade: what AA takes out between 18 and 20 kHz was on the disc. It doesn't cost attenuation above fs/2 the way a slow DAC filter does, since its own transition is a few hundred hertz and the main filter still does the anti-imaging. But it's still giving up real top octave to get rid of the pre-ringing, and whether that's worth it is taste. That's why it's a switch.

 

[Erogen]

Adele - 19 (2619s - 12 Track)
Intel Ultra 7 256 +GPU RTX 5060TI 32GB / DDR5 5600
AE Ver.1.32 [AE · 44.1k→176.4k · Kaiser 30M · f64 · DC·ISP·SUB20·PFR·HP]

Only Intel Ultra 7 265 - 6.60x
Intel Ultra 7 +GPU RTX 5060Ti
1 x AU - 33,15x
2 x AU - (6/6) - 35.39x
2 x AU - (7/5) - 39.09x
2 x AU - (8/4) - 40.29x

A second copy of the program is launched 3–4 seconds after the first one starts.
The disc has a duration of 43m 39s, and the program processes the data in 1m 05s.
Very fast (A good graphics card is essential here.)

 

[Erogen]

Continuation of various experiments.
Comparison:

Original recording

Recording version [AE · 44.1k→176.4k · Kaiser 30M · f64 · ISP·SUB20·PFR·HP·aHP]

Recording version – RMS -15 dB. High-pass filter 25Hz/-12dB

My ears tell me that this version plays the least aggressively.

 

[Erogen]

Original.

[AE · 44.1k→176.4k · Kaiser 30M · f64 · ISP·SUB20·PFR·HP·aHP]

RMS -15 dB. High-pass filter 25Hz/-12dB