Friday 13 September 2013

MEASUREMENTS: PCM to DSD Upsampling Effects (JRiver MC19 Beta).

We're continuing to see a push into the DSD domain with renewed talk of music release as digital downloads requiring the purchase of a DSD DAC to natively play (eg. recent Acoustic Sounds DSD releases). For now, I have already voiced some concerns about DSD including practical issues like the gross limitations of the file format itself. I demonstrated the noise characteristics for both DSD64 and DSD128 in my TEAC measurements. Furthermore, I have shown that there already exists many SACD's appearing essentially to be upsampled PCM from standard sampling rates of 44/48 kHz (remember, it's almost a given that any music we listen to created in the last 20 years has at some point been through a PCM stage except specified pure analogue recordings or those specifically recorded in DSD with minimal processing).

Some writers have voiced that even the process of upsampling PCM to DSD will imbue the music with some of DSD's beneficial properties, is this true? If so, what happens?

Well, thanks to ongoing advancement in the computer audio world, we can now easily have a way to listen to our PCM music as a converted DSD stream... Enter JRiver Media Center 19 and it's ability to stream PCM --> DSD64/128 in realtime to a compatible DAC. [Note that this should also be possible using ASIOProxy from foobar - not tried personally.] This will allow an easy way for everyone to listen for themselves what happens to the sound either as the original PCM or transcoded to DSD with the assurance that we're comparing "apples-to-apples" with the same mastering.

First, as has been my custom, let's start with some objective measurements to see what the DSD encoding does to test signals.

I. Objective Measurements

General Setup:
AMD A10-5800K HTPC Win8 x64 running JRiver 19.0.37 (ASIOProxy workaround for DSD128) -> shielded USB -> TEAC UD-501 -> shielded 6' RCA -> E-MU 0404USB -> shielded USB -> Win8 laptop

TEAC DAC settings:
     - PCM: "SHARP" BB PCM1795 filter, no upsampling (default)
     - DSD: FIR3 analogue filter (closest volume match to PCM output)

JRiver Media Center 19.0.37 setting:
     - ASIO buffer set to "minimum hardware size" since someone suggested it sounded better :-) - no stuttering encountered playing music.

First, let's have a look at the 1kHz SQUARE WAVE off the digital oscilloscope (24/44 source):

PCM:

DSD64 realtime conversion:

DSD128 realtime conversion:

As you can see, volume is about the same with the DSD FIR3 filter on the TEAC vs. PCM SHARP; both output ~2.85V peak with the square wave. What is also very obvious is how clean the PCM is vs. DSD. Notice the extra high frequency noise for both the DSD64 and DSD128 traces, with the DSD128 clearly less noisy. No surprise, right? If you've looked at the objective results from DSD here and elsewhere, this is pretty "normal" for DSD.

IMPULSE RESPONSE (16/44 PCM impulse):
PCM SHARP filter:

DSD64 realtime upsampling:

DSD128 realtime upsampling:

Noise is again very evident in this "zoomed in" impulse response measured at 24/192 especially from the DSD64 process. Although impulse response graphs can be excellent with MHz sampling rate (this is often a "talking point" in the DSD/Sony ad literature over the years), when resampling PCM to DSD, we're still hampered by the PCM signal's original sampling rate (eg. 44kHz). Evidently JRiver uses a typical linear phase reconstruction filter; hence the symmetrical pre- and post-ringing.

DUNN J-TEST:
PCM:
16-bit
24-bit

DSD64 realtime upsampling:
16-bit
24-bit
DSD128 realtime upsampling:
16-bit
24-bit
No meaningful differences between PCM and the DSD realtime conversion. This also means no evidence of worsened jitter with all that extra processing converting PCM to DSD in the computer at least based on the spectral output of this test (typically, this computer's AMD CPU utilization went from <5% with PCM to ~15% for DSD upsampling). DSD64 is obviously of enough resolution to accurately demonstrate the jitter modulation signal in the LSB for the 16-bit test.

RightMark:

Calculations done in the AUDIBLE SPECTRUM (20-20kHz).
Frequency Response
Noise
THD
These graphs of upsampled 24/192 test tones echo the results in the TEAC DSD Measurements back in May. I used KORG Audiogate to convert PCM to DSD back in May and it looks like the mathematical process in both JRiver and Audiogate are of similar precision. Within the audible spectrum, the PCM and DSD measurements are all very similar. A good indication of high precision in the conversion process. What is again evident is the noise once we get above 20kHz especially with DSD64.

II. Subjective Experience

As per the premise of this blog being "more objective", I'm not going to write pages on that which is experienced for oneself. However, I'll put down a few thoughts for consideration...

Listening gear:
     - Headphones: Sennheiser HD800 off TEAC DAC
     - Speaker system: TEAC UD-501 --> Simaudio Moon i3.3 --> Paradigm Signature S8 v3 (standard OFC cables)

The DSD conversion process through this TEAC DAC does change the electrical output as seen by the objective measurements above. This alone means that it's real compared to the identical measurements found with different bitperfect software and digital cables previously reported.

There does seem to be a change in the perceived detail of the sound subjectively through the gear I listed... Note that I'm taking the liberty here to not subject myself to a blind test so I fully admit that I could be wrong on this :-). Furthermore the fact that since it's not an instantaneous 'flip', echoic memory is prone to be unreliable. With these caveats, my current feeling is that both DSD64 and DSD128 conversion adds a potentially euphonic characteristic to the sound. No, IMO, it's not a dramatic difference when listening volume is controlled. [For those using the TEAC DAC, remember that the default FIR2 filter for DSD is louder than PCM by ~2.5dB - this could of course be misconstrued as sounding "better" for DSD.]

What do I hear? As I mentioned in my previous post on getting DSD128 upsampling working on the TEAC, I think the sound is less "etched". There's a pleasant subtle added smoothness to the transients. I think many may describe this as being less fatiguing, maybe less of the "digital glare". I couldn't specifically put a preference on DSD64 vs. DSD128 but knowing the ultrasonic effects, it wouldn't take much to convince me that DSD128 is better since the ultrasonic noise is further away from the audible spectrum. However, if you believe that the noise itself creates euphonia, it's also conceivable that DSD128 would sound closer to PCM than DSD64. Maybe.

I listened to a few standard 16/44 albums in DSD128 like a first pressing Michael Jackson Thriller, the well recorded Al Di Meola Winter Nights, and Suzanne Vega's Solitude Standing. They all sounded great. Like I said, marginally smoother than PCM. I think poorly recorded harsh albums may benefit even more - for example Alan Silvestri's The Avengers score is mastered in "modern" overcompressed fashion with DR9 average dynamic range (not good for an instrumental soundtrack IMO). DSD128 upsampling seemed to make it more listenable for longer duration.

Vinyl rips (24/96) of Tracy Chapman's Fast Car and Whitney Houston's One Moment In Time sounded very nice as well... "Extra" analogue from digital from vinyl :-). Again, the inability to instantaneously switch between PCM-to-DSD makes it hard to A-B compare reliably.

Unfortunately I did not take a screenshot of the phase measurements, but it looked good. Listening to phase-effect tracks such as those encoded in Q-Sound like Def Leppard's Rock On (David Essex remake off Yeah!), and Roger Waters' Too Much Rope (off Amused To Death) nicely created the impression of spatial surround and depth. Whether that sense of depth is any better with the DSD upsampling is of course debatable.

III. Conclusion

1. PCM to DSD upconversion is a DSP process. The signal output is measurably different.

2. Noise shaping pushes the DSD quantization noise into the ultrasonic frequencies as expected. In DSD64 it rises above the noise floor almost right at 20kHz, and in DSD128 it starts around 40kHz. (I vote for pushing it up to 40kHz as less likely to cause distortion through the amp & speakers.)

3. Pre- and post-ringing is similar to standard PCM with upsampling using MC19's algorithm so this would not explain any audible differences.

4. The algorithm used in JRiver MC19 does a good job with maintaining classic measurement parameters like frequency response, dynamic range, and distortion from 20-20kHz  - basically this means the math is as expected and fits the DSD output profile. Results are similar to the KORG AudioGate software converting PCM-to-DSD.

I can't help but wonder if what's happening here is like tube amps and analogue playback (eg. vinyl). Objectively the DSD conversion adds distortion but the anomalies are not perceived as objectionable and in some material, the added noise and imprecision actually makes it sound less "sterile", "clinical", more "real" (conversely being in an anechoic chamber is disturbingly unreal due to the profound silence). It would make sense to me that some people could prefer DSD64 over DSD128 upconversion since DSD64 will give you more of that distortion. Even though the noise is ultrasonic in nature as measured off the DAC, nonlinearities in the playback system like your headphones and speakers (perhaps certain amps as well) could create audible intermodulation. Maybe for certain music, this could be especially beneficial.

Out of curiosity... For anyone out there with the EMM Labs DAC2X which upsamples to DSD128, it'd be great to have a look at what the measurements are from that unit! With all the positive press about how this DAC sounds (ahem... $15.5K), I have yet to see any measurements... I wonder what a 16/44 impulse response looks like for example to see if it bears similarity to what JRiver is doing. How about the ultrasonic noise with DSD64 & analogue filter strength? Does the EMM upsampling process for PCM result in similar frequency response pattern?

In any case, give this PCM-to-DSD process a try at some point when you can. If nothing else, at least to say you've experienced it... See if you can perceive a difference and/or judge if it's beneficial for yourself.

Tonight's music: Valery Gergiev & LSO's Mahler Symphony No. 9 (2011). Nice recent classical recording available in SACD format as well.

---------------------------------------------
You might notice that I turned on AdSense on the blog.

As I have said in the past, my intention for this blog has never been about making money. I have no formal relationship with any company so have no sales incentive and am not interested in making this some kind "publication" other than what it has been - a blog about my own journey in audiophilia with a bent towards finding answers using empirical/objective means. That remains my main interest.

Nonetheless, it's trivial to "flip" the AdSense switch. I would have no idea what Google tries to market to you, and trust that the layout won't be distracting (I've switched off some questionable types of ads like for dating sites or of a sexual nature). If it gets me a few bucks for my digital downloads for what I do as a hobby anyway, I'll be happy with that!

Best regards...

17 comments:

  1. Could you try doing loaded measurements, that is for example 30 ohms in parallel with the input?

    Maybe the high frequency noise does influence how the amp behaves driving an actual load?

    ReplyDelete
    Replies
    1. I hope to measure what th output from the amp will look like.

      For now, all I'm demonstrating is at the level of the line level output from the DSD DAC.

      Delete
  2. Thanks for doing the measurements (I also admire your work on the other measurements of the TEAC UD-501).

    I am currently listening with 2xDSD (I have not tried 1xDSD), and using the FIR2 filter. My impressions mirror yours. That is, 2xDSD gives a smoother sound, cleaner treble, and appear to sound more real; a better illusion that the singer/band is in my living room rather than me listening to a recording. Of course, it is still far from the real thing, but 2xDSD seems to get another step closer. With 2xDSD, I also noticed a wider and deeper soundstage with better layering (more space between layers).

    I compared the filters FIR2 and FIR3. FIR3 seems to sound closer to regular PCM. FIR2 has more of the effects that I described earlier (smoother, cleaner, better soundstage) compared to FIR3. I tried to adjust my volume to compensate for the increased volume of FIR2, and still found that FIR2 sounds better than FIR3. Michael Lavorgna from Audiostream also seem to prefer FIR2 in his review of the TEAC UD-501, and in forums I read a few posts where people also prefer FIR2. Any chance you can measure FIR2 and compare that to FIR3 (like what you did with the SLOW and SHARP PCM filter in your previous posts)?

    My system:
    TEAC UD-501 --> Octave V70SE integrated amp --> Dynaudio Contour S3.4 speakers.
    ...--> PS Audio GCHA headphone amp --> Audeze LCD-2 headphones.
    Interconnects and speaker cables are Nordost Heimdall.
    Power cables are Shunyata. Power condition is Torus RM15.

    ReplyDelete
    Replies
    1. I also usually listen with FIR2 as well. It's really hard getting the A/B comparison right between PCM vs. FIR2 vs. FIR3. You could be right that FIR2 could sound different/better.

      Technically, the difference between FIR2 and FIR3 looks minimal. The Fc is 90kHz with FIR2 and 85kHz with FIR3, so the lowpass is just *slightly* at a lower frequency. The main difference is of course the gain differences with there being a 1.8dB difference with FIR2 louder.

      Delete
    2. I have been using FIR3 yesterday and I am liking it more and more. FIR3 is less forward and smoother than FIR2. With FIR3, the details are still there, just not as obvious. So initially, FIR2 might sound better, but over time FIR3 might be better.

      Because of the volume differences between FIR2 and FIR3, I think it is best to just listen to each on for a few days with various genres of music and then switch filter. I will stick with FIR3 for a few days and see how I feel about it.

      Delete
  3. > Objectively the DSD conversion adds distortion but the anomalies are not perceived as objectionable and in some material, the added noise and imprecision actually makes it sound less "sterile", "clinical", more "real" <

    RMAA says the distortion becomes less while using up-sampling thus there are no added harmonics (as in tubes or vinyl) only the addition of noise.
    What is obvious is the roll-off at 20kHz of -0.5dB.
    Some will say this is so little and who can hear up to 20kHz.
    In experiments (in my much younger years) I researched a JK filter for CD players. It consisted of a low pass 24dB/oct with transistors (no op-amps) used instead of the steep brick-wall fashionable in those days and was claimed to improve SQ considerably.
    After building it myself I could pick out which filter was in use (blind) and it bugged me why as both were 'flat' only a slight deviation at the 20kHz extreme.
    Was it the circuit, as claimed by the manufacturer, that give it a more 'real' sound ?
    It turned out a roll-off of -0.5dB at 20kHz gives the sound slightly more 'body' but no apparent loss of highs at all (one would expect a slightly less edgy sound).
    The same circuit tuned higher up made it sound like the original filter.
    It even worked with 2 filters behind each other turned out later.

    I would guess the roll-off seen in the DSD signal adds more 'euphony' than the lower amount of distortion (I cannot see where the distortion increases in the RMAA numbers).
    It would be interesting to see an ARTA test showing 1st to 5th harmonics over the entire FR.
    That would me more conclusive than just an RMAA number that isn't clear how it was created.

    Too bad the 'audio test' is not possible ?
    subjective found differences are easily 'confirmed' by others, proving its existence technically is the difficult part.

    the tests above only show an (not unexpected) increase in noise as far as I can deduct.

    It should be noted that all 1 bit (noise shaping) DAC's convert the incoming PCM 'bits' to a bitstream as well, only ladder DAC's output real 'steps' so in essence all those modern DACs already do the PCM to DSD conversion already.

    Of course MASH and similar 'construction' DAC's (PCM1795) seem to be part ladder, part noise shaping.

    Real blind tests with statistical relevance may be more conclusive about audible differences.
    They are easy to do but time consuming and take a long (tiring) time.

    ReplyDelete
    Replies
    1. A min. phase low pass filter even above 20 kHz can cause considerable phase shift way below 20 kHz...


      The only thing I see causing audible differences is the added ringing and noise with DSD.
      The added ringing will smear transients, probably cause aliasing, and generally reconstruct something which is different from what it should be.
      The high frequency noise may cause problems in downstream electronics and also the headphone drivers themselves.

      Delete
    2. Btw, I'm not saying that phase shift is necessarily audible with music, but that even tiny phase differences between the channels can cause clearly audible differences.

      Delete
    3. True - at the end of the day, controlled listening tests are very important!

      I've been playing with ARTA as well and will need to spend more time on that in the days ahead. The RM distortion numbers calculated from 20-20kHz suggests the signal is not being impacted at all/barely in the traditional audio frequencies. I'm sure if the software calculated from 20kHz-40kHz it would be terribly nasty.

      Delete
  4. I found this blog after a few days of becoming "hooked" on the "sound of DSD" conversions from any and all PCM to DSD. I think you are on to something.

    I am completely hypnotized by the change. I am planning on using DSD as my archival format for well recorded Vinyl Releases.

    [iMac, Korg AudioGate, DSD downloads, PS3 (DSD Disc + Analog out) and my iTunes Library]

    ReplyDelete
    Replies
    1. Congrats on finding good sound!

      Happy listening :-)

      Delete
  5. Wideband noise measurements of Teac Ud 501 in PCM and DSD modes:
    www.computeraudiophile.com/blogs/miska/playback-formats-teac-ud-501-ti-pcm1795-458/
    I found strange that in PCM mode output there is not an analog low-pass filter for the high frequency 6 bit ladder noise.

    ReplyDelete
  6. Miska's measurements show that it best to use PCM 'sharp' or DSD to get the best results from this DAC.

    The high amount of ultrasonic garbage that is put out by the DAC in all other filter settings may not become audible on well made amplifiers but can sure cause a 'bite' when amps are used behind it which may produce audible signals due to non-linear behaviour (interference and/or detection)

    It seems only mild analog filtering is used which seems optimised for the highest possible bandwidth. A manufacturers choice.
    Manufacturers are almost obligated to supply different digital filter topologies in order to be taken serious and boost sales.

    ReplyDelete
  7. Thank you for the great blog with so much effort of yours.
    The details you provided regarding the Teac DAC UD-501 are much helpful and supported my decision of upgrading my Asus Essence ST to the UD-501.
    My stereo chain is NAD C162 PRE > 2 x Mono Bridged NAD C272 AMPs and pair of Paradigm Studio 100 v.2.
    I have received the DAC recently and still learning the sound letting it to burn a bit.
    I thing that this post is outdated since the updated Teac ASIO 1.02 driver works great with the latest J.River 19.0.74 both for PCM and DSD with no ASIOProxy workaround.
    Yet there is something I missing regarding the J.River DSP output encoding of PCM to DSD.
    While I use the 2xDSD in DOP format for the PCM records the while checking the output Audio Path shows the following
    Input : 44.1kHz 16bit 2ch from source ...
    Changes : Encode as 2xDSD in DoP format
    Output : 352.8 kHz 32bit 2ch using ASIO (direct connection)
    Teac readings are all good as well : DSD 5.6MHz / DoP

    However, while I change the DSP Output encoding to 2xDSD in native format, which I believe is a preffered way to transport the stream to the UD-501, J.River produces some kind of error message in the Audio Path -
    Input : 44.1kHz 16bit 2ch from source ...
    Changes : Encode as 2xDSD
    Output : DSD 5.6MHz 1bit 2ch using ASIO (not using enough bits to output the input directly)
    The DAC readings are still good DSD: 5.6Mhz native.

    The DSD records bit-streaming works well both for x64 2.8MHz and x128 5.6MHz - showing in the J.River audio path DSD 2.8MHz/5.6MHz 1bit 2ch using ASIO (bitstreaming) .

    Do you understand what does "not using enough bits to output the input directly" means ?
    I had no much time to compare the sound quality difference between the DoP vs ntive 2xDSD modes, but Is there any sound quality impact ? Or might it be a software glitch ?

    ReplyDelete
    Replies
    1. Here is the answer I have received from the MC beta team :
      "Native DSD requires ASIO, but ASIO expects 32-bits but native DSD is 1-bit. I think this is why MC displays that message but it can be safely ignored."

      Delete
  8. Thanks for the hard work you put into this investigation. I came across your post when considering whether to listen to the Yes High Vibration Box set - it seemed to me inherently silly to convert 96Khz 24-bit PCM to DSD for the SACDs but your post has assured me that I should give it a listen. I do love Pure DSD, however, especially the many great discs that Telarc put out in the first decade of the 21st century.

    ReplyDelete