Audibility thresholds of amp and DAC measurements

[pkane]

OK, I suppose an excerpt would suffice (in view of copyright) rather than the entire track?

For DeltaWave, a 1.5 - 2 minutes clip is sufficient for analysis

 

[xr100]

OK--didn't want to use the same track that I used earlier, was thinking about what to use and... the answer became obvious.

For the avoidance of any doubt over musical taste, anyone downloading and listening to these files will be "Rickroll'd"--LOL. First 90 seconds of the 7" Instrumental version.

Files (7zip format containing FLACs.)

Link expires in 7 days, so get your skates on...

As lossless was requested and export to floating-point FLAC is not available, to prevent digital overs, above all with the all-pass filter, all files are (post-main process) reduced in level by -6dB and dithered to 24-bit. (Source is 16-bit/44.1kHz.) That is the only processing to the "original."

 

[pkane]

OK--didn't want to use the same track that I used earlier, was thinking about what to use and... the answer became obvious.

For the avoidance of any doubt over musical taste, anyone downloading and listening to these files will be "Rickroll'd"--LOL. First 90 seconds of the 7" Instrumental.

Files (7zip format containing FLACs.)

Link expires in 7 days, so get your skates on...

As lossless was requested and export to floating-point FLAC is not available, to prevent digital overs, above all with the all-pass filter, all files are reduced in level by -6dB and dithered to 24-bit. (Source is 16-bit/44.1kHz.)

Got it. So, let's do these one at a time.

First, the soft-clip file. Correcting for nonlinear phase differences, I get this (turning on non-linear EQ settings in DeltaWave). Check the results on the bottom status bar, too, as these are additional metrics computed by DW:

Now, the Soft-Tube Trident A result:

And the 8th-order all-pass filter (also using alternate clock drift computation):

DeltaWave settings used for all of the above are shown here:

 

[xr100]

Got it. So, let's do these one at a time.

DeltaWave settings used for all of the above are shown here:

Thank you very much; I've tried your settings and, just to be sure, got the exact same results.

I've created some more test files (same 90 seconds "Rickroll'd" instrumental version reference.)

Waves SSL E Channel (could use G, too; identical plug-in, same except for the EQ's gain/Q interaction) "emulation" clearly has a basic signal-path model*--the only significant distortion components are the 2nd and 3rd harmonic. (Just under -60dB for the 3rd harmonic and just under -90dB for the 2nd harmonic with a 1kHz sine input at 0dBfs.) I bypassed the EQ and dynamics sections, and it has a flat frequency response in this mode. (The "Analog" option was also turned off, as all that does is add a small amount of noise.) A caveat is that it does not internally upsample, so significant aliasing can be generated when running at 44.1/48kHz.

(*Not that this is a bad thing for creative use in production, rather than throwing the "kitchen sink" of colouration processes at the signal, albeit it doesn't say what it does on the "tin!")

Essentially, it's rather subtle distortion, so I thought it would be useful to compare:

Testing a couple of non-linear processes that certainly have "memory," here's the Waves L316 16-band linear phase multiband "brickwall" limiter. Default settings used, except "Threshold" and "Margin" set to -1dB.)

The compressor from the Sonnox Oxford Dynamics plug-in (DSP was ported from the Sony OXF-R3 console) -- "Compressor" section only, default settings, except "Threshold" set to -3dB:

And finally, WaveArts Tube Saturator Classic. This is said to use a "circuit simulation" (i.e. a kind of real-time "SPICE" simulation) of a circuit that contains the 12AX7 valve. (The "Classic" version is free but uses more CPU than the paid for version, "Tube Saturator 2," which has some additional features including an "Oversampling" option that the "Classic" version lacks.)

Default settings were used with "EQ" and "FAT" modes off, which means that "DRIVE" was set to the middle position. The signal was scaled so that it was reduced by -18dB before input to the plug-in, then raised by 18dB following the plug-in.

Almost exactly the same calculated "DF Metric" value as the Waves L316 16-band multi-band limiter... (and almost 20dB worse than the 8th-order all-pass filter.) Hmm...

 

[xr100]

Another couple of tests using the same reference file.

TPDF Dither (dither ONLY--no noise shaping loop) to 8-bits -- "DF Metric" = -31.6dB
Truncate to 8-bits -- "DF Metric" = -35.4dB

Hmm... well, we don't know the noise floor of the reference, so how about 4-bits?

TPDF Dither to 4-bits -- "DF Metric" = -12.8dB
Truncate to 4-bits -- "DF Metric" = -15dB

Listening to the "difference" between the source and comparison files, the dithered 4-bit sounds mostly like white noise with, say, some HF percussion breaking through; the truncated 4-bit sounds like a mess of crackly quantization noise.

And, just in case I have my wires crossed in relation to what is a "better" "DF Metric" value... comparing the reference file to itself -- "DF Metric" = -300dB.

 

[pkane]

Another couple of tests using the same reference file.

TPDF Dither to 8-bits -- "DF Metric" = -31.6dB
Truncate to 8-bits -- "DF Metric" = -35.4dB

Hmm...

And, just in case I have my wires crossed in relation to what is a "better" "DR Metric" value... comparing the reference file to itself -- "DR Metric" = -300dB.

-300dB is the lower cutoff point in DeltaWave to eliminate negative infinity, just in case someone claims that they can hear the difference of -300dB

 

[xr100]

-300dB is the lower cutoff point in DeltaWave to eliminate negative infinity, just in case someone claims that they can hear the difference of -300dB

How about -299dB? ;-)

 

[pkane]

How about -299dB? ;-)

That might matter, but only if your name is Rob Watts of Chord Electronics who says this:

My own tentative conclusions (or rule of thumb) are that one can hear levels of noise floor modulation down to -200dB - currently we can measure noise floor modulation at -180 dB

 

[pkane]

Another couple of tests using the same reference file.

TPDF Dither (dither ONLY--no noise shaping loop) to 8-bits -- "DF Metric" = -31.6dB
Truncate to 8-bits -- "DF Metric" = -35.4dB

Hmm... well, we don't know the noise floor of the reference, so how about 4-bits?

TPDF Dither to 4-bits -- "DF Metric" = -12.8dB
Truncate to 4-bits -- "DF Metric" = -15dB

Listening to the "difference" between the source and comparison files, the dithered 4-bit sounds mostly like white noise with, say, some HF percussion breaking through; the truncated 4-bit sounds like a mess of crackly quantization noise.

And, just in case I have my wires crossed in relation to what is a "better" "DF Metric" value... comparing the reference file to itself -- "DF Metric" = -300dB.

So TPDF dither actually makes the DF metric worse by a few dB? How about trying a shaped dither?

 

[xr100]

That might matter, but only if your name is Rob Watts of Chord Electronics who says this:

Hey, if he can develop at least a 6-channel DAC (so PC-based digital loudspeaker management can be used) with ultra-high performance for a non-insane price, I'll buy it. (Where my idea of "non-insane" in relation to audio is closer to what Paul Messenger, writing for Hi-Fi Choice magazine in those days, used to [irritatingly] call "beer budget" rather than Chord Electronics. Someone wrote in asking how much beer he drank!)

 

[xr100]

So TPDF dither actually makes the DF metric worse by a few dB? How about trying a shaped dither?

Probably worse--more noise albeit pushed out of audibility's way into less sensitive frequencies.

I used "JS - Bit/Reduction w/ Noise Shaping" in Reaper--the noise shaping loop would end up overloading the output (which results in the noise shaping automatically being turned off!) if I'd tried to turn it on. So, I'll switch to Waves L1+ Ultramaximizer -- "Dither" set to "Type 1," "Shaping" set to "Ultra." All other settings at default--"Domain" was set to "Digital" to ensure that it was working with the sample values rather than internally upsampling the sidechain signal to work with an estimated "reconstructed" waveform (as in the "Analog" setting for the old version that I have, or the "True Peak" compliant setting that is available in the new "25th Anniversary" version.) Hence, there was no action by the limiting section. Bit-depth set to 8-bits: "DF Metric" = -29.3dB.

Almost forgot--I have a modified version of the "JS - Bit/Reduction w/ Noise Shaping" plug-in for which I changed the script to turn back on noise shaping in case of overs. In which case, FWIW, "TPDF Dither" selected with "Noise Shaping" on, bit-depth of 8-bits: "DF Metric" = -26.7dB.

 

[j_j]

For distortion, (2).

To quote from Douglas Self's "Audio Power Amplifier Design Handbook":

"Phase and group delay have been an area of dispute for a long time. As Stanley Lipshitz et al. have pointed out, these effects are obviously perceptible if they are gross enough; if an amplifier was so heroically misconceived as to produce the top half of the audio spectrum 3 hours after the bottom, there would be no room for argument. In more practical terms, concern about phase problems has centered on loudspeakers and their crossovers, as this would seem to be the only place where a phase shift might exist without an accompanying frequency-response change to make it obvious. […] This controversy is of limited importance to amplifier designers, as it would take spectacular incompetence to produce a circuit that included an accidental all-pass filter. Without such, the phase response of an amplifier is completely defined by its frequency response, and vice versa; in Control Theory this is Bode’s Second Law, and it should be much more widely known in the hi-fi world than it is. A properly designed amplifier has its response roll-off points not too far outside the audio band, and these will have accompanying phase shifts; there is no evidence that these are perceptible."

(Slight caveat--Self's book is not a definitive treatise on questions of audibility.)

More to the point, it's not "post analog era". Things that are barely, if at all, possible in analog are easy in digital (and vice versa). Codecs (Yes, I know, not terribly hi-fi) are a classical example.

Phase shift of more than 15 degrees or so across an ERB (or critical band) will provoke sensation differences when the right signal comes along. That's a reasonably decent estimate. At very low frequencies, even less may be the case, but I'm unaware of anyone actually creating a signal to demonstrate that.

 

[j_j]

Sound is a waveform (in time domain).

At one point in a space, 4 waveforms measuring different things. Please don't stick to pressure, that confuses everything.

 

[bennetng]

Got it. So, let's do these one at a time.

First, the soft-clip file. Correcting for nonlinear phase differences, I get this (turning on non-linear EQ settings in DeltaWave). Check the results on the bottom status bar, too, as these are additional metrics computed by DW:
View attachment 46671

Now, the Soft-Tube Trident A result:
View attachment 46672

And the 8th-order all-pass filter (also using alternate clock drift computation):
View attachment 46673

DeltaWave settings used for all of the above are shown here:
View attachment 46674

Does it mean DeltaWave users should always use settings that yield the best possible nulls? Even when different file pairs require different settings?

This point is important if you want to do something similar to Gearslutz's tests, and especially important for people who only look at numbers without downloading and examining the recorded audio files.

 

[Serge Smirnoff]

Here is my results for @xr100 files:

8th Order All-Pass.flac_cut.wav(44)__Original (-6dB).flac(44)__mono_400-11.9671-6.9672-0.0176

Soft Clip.flac(44)__Original (-6dB).flac(44)__mono_400-28.9115-23.7135-0.1949

Softube Trident A-Range.flac_cut.wav(44)__Original (-6dB).flac(44)__mono_400-29.3091-25.0161-0.1366

Similarity of artifact signatures:

Conclusions:

(1) Soft Clip and Softube samples have very similar artifact signatures (0.18dB, while 1.5-2.0dB is a critical distance for a good relation of df levels to subjective scores), so Softube sample will sound more close to the original than Soft Clip but the difference is really subtle (needs very careful listening).

(2) 8allpass sample has different artifact signature at the distance of ~1.44dB from other samples. Such distance is critical but taking into account that 8allpass sample has substantially higher/worse Df level (-6.97dB in comparison to -25.02dB for Softube) it is highly likely that this distortion will be more audible despite the distinction in signatures.

In other words 8allpass sample will sound less close to the original than two others; the SoftClip and Softube samples will be hard to discern from each other by listening.

This is correct way of using df-metric.

 

[xr100]

Here is my results for @xr100 files:

Thank you. Would be interested to see your analysis of the other files that I tested:

Download

Once again, link expires in 7 days.

 

[Krunok]

Phase shift of more than 15 degrees or so across an ERB (or critican band) will provoke sensation differences when the right signal comes along. That's a reasonably decent estimate.

Interesting. Has that been demonstrated?

 

[pkane]

Does it mean DeltaWave users should always use settings that yield the best possible nulls? Even when different file pairs require different settings?

This point is important if you want to do something similar to Gearslutz's tests, and especially important for people who only look at numbers without downloading and examining the recorded audio files.

If you want to duplicate Gearslutz tests, you'll want to turn off non-linear EQ settings in DeltaWave. These correct for more than the simple clock drift and linear gain differences -- these correct for the non-linear effects, as well, which are often the result of filters used by DACs and ADCs.

To be fair to Serge, the DF metric is defined as using only linear corrections such as a linear sampling rate difference or a linear difference in amplitude. So, to view the same DF metric that he describes, it's best to also turn off non-linear EQ in DeltaWave.

Here's an example of what DeltaWave non-linear EQ corrects that both, DF-metric and Gearslutz metrics do not. This is the phase difference between the original and the corrupted file (this is with Softube Trident comparison @xr100 uploaded yesterday):

If you believe such a phase difference is significant and audible (you can test for this in DeltaWave) then don't correct for it in DeltaWave - turn off non-linear EQ. I usually check to see how large the phase differences are and at what frequencies. Something like this may be very audible (the 8th order all-pass filter from @xr100):

 

[j_j]

Interesting. Has that been demonstrated?

Well, yes, how do you think I know that?

 

[xr100]

Conclusions:

(1) Soft Clip and Softube samples have very similar artifact signatures (0.18dB, while 1.5-2.0dB is a critical distance for a good relation of df levels to subjective scores), so Softube sample will sound more close to the original than Soft Clip but the difference is really subtle (needs very careful listening).

[snip]

In other words 8allpass sample will sound less close to the original than two others; the SoftClip and Softube samples will be hard to discern from each other by listening.

See previous post for an overview of the behaviour of "Softube Trident A-Range" and the soft clipping process.

Below is a test using a ramped signal (up to 0dBFS over 15 seconds) vs. output. The static waveshaping activity of the "soft clipper" can be seen. By comparison, the "Softube" seems to barely diverge, although, as mentioned, its internal behaviour (aka "modelling" of the hardware's signal path) is unknown but presumed not to be a single static waveshaper. Another caveat is that the "Softube" generates even order harmonics, so one would expect an asymmetrical result and the test signal was positive only.

Underneath are spectral plots for both "Softube" and the "soft clipper," 1kHz sine input, at 0dBFS, -10dBFS and -60dBFS. (The oversampling function in the "soft clipper" was turned off.)

The "soft clipper" basically passes the signal untouched at -10dBFS, as would be expected since its soft clipping action starts at ~-4dBFS, and so there was no need to include the -60dBFS plot. The "Softube" plot show harmonic distortion at -60dBFS.

As a symmetrical waveshaper, the soft clip process plots only show odd order harmonics; the Softube plots also show even order harmonics--although note that that at -60dBFS, only odd order harmonics can be seen.

(N.B. 44.1kHz sample rate was used, except for the spectral plots which were taken at 96kHz.)

It can be seen that the behaviour of the two processes is very different.