Review and Measurements of iFi iPurifier S/PDIF Digital Audio Filter

[DonH56]

There are several threads about jitter in the Audio Reference Library sub-forum that might help answer some of the questions about jitter types and audibility.

Note that unless the recovered S/PDIF clock is used by the DAC to generates its output signal then using a PLL to following the wandering source clock does not necessarily cause jitter in the analog(ue) output.

Using the two most basic and broad jitter definitions, random jitter (RJ) will raise the noise floor (reducing SNR) and deterministic jitter (DJ) adds distortion spurs. DJ is generally much (MUCH!) more audible and spurs can be non-harmonic since they are related to the signal and clock and/or whatever is causing them (e.g. power supply or other EMI/RFI).

HTH - Don

 

[obo]

The "eye pattern" measurement appears more relevant than the no-signal jitter (raise and fall improves by 10-15%)? Shouldn't this raise/fall be measured at the output of the DAC? One might take a real signal, convert through ADC, apply to either 1) ifi+DAC and 2) DAC, and compare?

Nice measurements though.

 

[amirm]

The "eye pattern" measurement appears more relevant than the no-signal jitter (raise and fall improves by 10-15%)? Shouldn't this raise/fall be measured at the output of the DAC?

First, welcome aboard .

On your question, sharp transition signals only exist in the S/PDIF domain (bandwidth of 100+ Megahertz). The output of the DAC (bandwidth of tens of Kilohertz) cannot remotely reproduce anything close to it. In the J-Test measurements, what is there is a 12 Khz square wave for example. But all of its harmonics are truncated so what we get is just a sine wave at 12 Khz.

Is this what you were asking?

 

[amirm]

How 'bad' does jitter have to be for humans to hear it?
How does it translate to distortion in the analogue domain?

Answering the second part, that depends on the spectrum of jitter. If Jitter is sinusoidal, then you get two sidebands, one above and one below, at the frequency of our main ton plus/minus jitter frequency. See the measurement I did with 4 Khz induced jitter. How tall it is depends in the excursions (amplitude) of the jitter.

Once we have that spectrum as I have been showing, then we need to apply two things to it:
1) Masking thresholds
2) Threshold of hearing

#1 says that the lower the frequency of jitter, the less audible it is. #2 says it better be higher than about -120 db at mid-frequencies to matter.

The late Julian Dunn (creator of J-Test and much work in digital audio) has this nice graph showing effect of masking on jitter. The db numbers I show can be converted to time values (Y axis) and then compared to this graph:

As a general rule, I don't like to see any jitter spikes rise up above -96 db or so.

 

[DonH56]

Amir, where is that from?
Thanks - Don

 

[obo]

Hi Amir,

Thank you for the follow-up. The graph seems right - smearing of lower frequency should be less apparent than that of a higher frequency.

When considering jitter effects, I propose that we should be looking for changes in shape of the time-evolved decoded signal. Within the given bandwidth of the DAC, that is. The 12kHz square might be too far from the capability of the audio decoder (the distortion from square too large already vs. jitter effect). However, a real audio signal is exactly within DAC capability (bandwidth) which is why I propose that a real signal be converted to digital, then applied to either ifi+DAC or DAC alone, and resulting output be compared to source. Perhaps there are other issues that might prevent a valid conclusion with such test?

(Or, such test might not be easily accessible within the J-Test?)

 

[bennetng]

As a sequel of my previous post about jitter measurement without analog conversion in this thread...
https://www.audiosciencereview.com/forum/index.php?threads/digital-distortion.2059/#post-55543

The attached files are measurements with analog conversion. RMAA html report and J-Test audio files included so that members can analyze them with any analyzers and any desired parameters.

The test was done by playing a 48k J-Test file using my Realtek ALC892, then use toslink to send the output to my SC-D70, and finally use my X-Fi to record SC-D70's analog output. Both the ALC892 and X-Fi are clocked at 48k, the only variable is SC-D70's operating rate (44 vs 48k). As shown in the RMAA report and J-Test audio files, the ones with non-matching rate (44k) really has less measured jitter, even in analog output.

The test is aimed at showing how SPDIF performs in some poor (if not poorest) conditions:

-SPDIF output: Realtek ALC892 on a $62 ASRock mainboard
-internal PCIE soundcard (ADC): X-Fi
-ancient external interface (DAC): Roland SC-D70 in 2001
-optical cable: allegedly more jittery than coaxial

In fact the 50Hz + harmonics are the ugliest things in the whole test, rather than jitter.

The original J-Test signal is attached in this post:
https://www.audiosciencereview.com/...eview-schiit-modi-2-99.1649/page-8#post-57415

 

[bennetng]

Pictures showing the clumsy and ugly connections.

 

[Jakob1863]

Amir, where is that from?
Thanks - Don

That is based on the calculations done by Julian Dunn in:

Julian Dunn—‘Considerations for Interfacing Digital Audio
Equipment to the Standards AES3, AES5, AES11’ Published in
‘Images of Audio’, the Proceedings of the 10th International AES
Conference, London, September 1991. pp 115-126.

Dunn used publications from Zwicker/Zwicker and Blesser/Kates for the calculations considering masking theory regarding masking of sidebands (of the high level tones) by this high-level tones and derived some graphs with worst case assumptions for various sampling systems.

After the publication from Benjamin/Gannon Dunn wrote :

"In the view of the more recent research, this may be considered to be overcautious.
However, the consideration that sampling jitter below 100 Hz will probably be less audible
by a factor of more than 40 dB when compared with jitter above 500 Hz is useful
when determining the likely relative significance of low- and high-frequency sampling
jitter."

(Julian Dunn, Measurement Techniques for Digital Audio, Audio Precision, Application Note #5, page 34)

 

[Jakob1863]

First, welcome aboard .

On your question, sharp transition signals only exist in the S/PDIF domain (bandwidth of 100+ Megahertz). The output of the DAC (bandwidth of tens of Kilohertz) cannot remotely reproduce anything close to it. In the J-Test measurements, what is there is a 12 Khz square wave for example. But all of its harmonics are truncated so what we get is just a sine wave at 12 Khz.

Is this what you were asking?

The J-test signal (introduced by the aforementioned Julian Dunn) originally consists of an undithered high level component at Fs/4 (in the digital domain it looks like a square wave at - 6dB, although it represents a Fs/4 sine wave at - 3dB) added to the second component, an undithered square wave of switching between "0" and 1 LSB at a low repetition rate (usually rate of Fs/192).

Which means in fact that a signal consisting just of the high level component isn´t a "J-test signal" .

 

[bennetng]

The J-Test signal I mentioned in my previous post is supposed to be the "official" one. Related discussion here:

https://hydrogenaud.io/index.php/topic,108928.0.html

The file I attached in ASR is extracted from Arnold's attachment.

I used this signal in various tests and the resulted jitter spectrum of using a simple 1/4 sample rate tone and the official one are in fact very similar or indistinguishable (in 24-bit... 16-bit J-Test signal will make some differences though). The only thing to worry about is the file's amplitude, as difference in level (dBFS) of the test signal will affect overall noise level proportionally.

 

[amirm]

Amir, where is that from?
Thanks - Don

It is from a private powerpoint presentation from Prism Sound (where Dunn was the CTO for a while). There is a more crude version of it in Dunn's AES UK paper:
JITTER AND DIGITAL AUDIO
PERFORMANCE MEASUREMENTS
JULIAN DUNN
Prism Sound
Cambridge, U.K.

 

[DonH56]

Thanks Amir, and Jakob.

 

[amirm]

The J-Test signal I mentioned in my previous post is supposed to be the "official" one. Related discussion here:

https://hydrogenaud.io/index.php/topic,108928.0.html

The file I attached in ASR is extracted from Arnold's attachment.

I used this signal in various tests and the resulted jitter spectrum of using a simple 1/4 sample rate tone and the official one are in fact very similar or indistinguishable (in 24-bit... 16-bit J-Test signal will make some differences though). The only thing to worry about is the file's amplitude, as difference in level (dBFS) of the test signal will affect overall noise level proportionally.

I am not sure it is the right test file. This is its spectrum:

Here is the spectrum of mine (created from C program):

When I have time I will pull them both into matlab to verify and report back.

 

[bennetng]

I am not sure it is the right test file. This is its spectrum:

View attachment 9919


Here is the spectrum of mine (created from C program):

View attachment 9920

When I have time I will pull them both into matlab to verify and report back.

Which file you are referring to? The files I posted in that HA thread are the incorrect ones, the supposedly correct files are the later ones in the same thread, in both 16 and 24-bit.

 

[amirm]

Which file you are referring to? The files I posted in that HA thread are the incorrect ones, the supposedly correct files are the later ones in the same thread, in both 16 and 24-bit.

I took them out of the zip file you post here.

 

[bennetng]

I took them out of the zip file you post here.

This one?
https://www.audiosciencereview.com/...f-digital-audio-filter.2189/page-2#post-59476
No. They are the recorded files, not the test signal.

 

[RayDunzl]

Maybe the Reference Forum could use a post with proper 16 and 24 bit J-Test WAV files...

 

[bennetng]

I realize different members use different analyzers (different versions of Audition, WaveSpectra, AP etc) with different FFT sizes, Window types and other parameters. A static screenshot is rather limited so I posted the actual recorded files. In this way other members can obtain more information from the recordings, and compare them with their recordings with the software they use, as well as making appropriate adjustments (e.g. volume normalization)

I originally wanted to post two 10 seconds snippets but they are over 2MB, so I trimmed them down to 8 seconds.

The original J-Test files are very compressible with non-audio algorithms like zip or 7z since they are just some repeated sequences without any dither.

Anyway, here is my analysis (24-bit only)
red: j48.flac
green: j44.flac

...and the zoomed version.

 

[amirm]

This one?
https://www.audiosciencereview.com/...f-digital-audio-filter.2189/page-2#post-59476
No. They are the recorded files, not the test signal.

Yes, that is the one.