• WANTED: Happy members who like to discuss audio and other topics related to our interest. Desire to learn and share knowledge of science required. There are many reviews of audio hardware and expert members to help answer your questions. Click here to have your audio equipment measured for free!

Evaluation of SBM Direct downconversion from DSD64 via 64 bit FIR filter 32k taps

Saidera

Senior Member
Joined
Jul 18, 2020
Messages
388
Likes
309
Location
Adelaide, South Australia, Australia
My primary interest is in DSP like SBM Direct because DSD to me is overkill. If the ‘best sounding’ PCM CD is made from DSD via SBM Direct, like Sony HK liked to assert, then I want to test that claim properly.

So how'd one go about doing that?

Like this perhaps:
https://www.audiosciencereview.com/...ound-quality-of-dsd.14773/page-20#post-517365
https://www.audiosciencereview.com/...ound-quality-of-dsd.14773/page-21#post-517534
https://www.audiosciencereview.com/...ound-quality-of-dsd.14773/page-22#post-518757

By getting SSMS DSD Mode, unless you have access to K-1327 in a mastering studio somewhere, or a Sonoma software system.

I would like some assistance in this matter.

It may not be possible to objectively test this claim that DSD CDs are better. Even mastering studios will regularly subjectively compare different ways of downsampling DSD, and SBM Direct is not always the first choice.

However Eiichi Ohtaki used SBM Direct for his 2011 remaster of his album A Long vacation (now in 360 Reality Audio, so it is not an unpopular album in Japan). Then he indirectly publicised it via book interviews etc although he had no interest in suffering the difficulties of creating music as opposed to merely listening, nor was he interested in using DSD or Blu-Ray to create anything or as a consumer.

I am still researching, but SBM Direct can apparently retain enough of the benefits of pure recorded DSD audio (if any) within the 22 kHz band it is limited to playing with thanks also to the proprietary curve which is different to the 1990s SBM's one, however details are kept hidden by Sony. I want to objectively verify this statement or to deny it and to accept that there are better dithers now. See http://archimago.blogspot.com/2020/07/summer-musings-post-hi-res-audio-why-hi.html#more
where the SBM process is shown to keep the noise floor down to around -140dB for most of the audible range and rises from 14kHz. How does SBM Direct compare?

You may have heard SBM Direct PCM if you bought SACDs dual layer (playable in CD player) or have DSD CDs from Hong Kong, or Yo Yo Ma’s 2009 Box set (full SBM Direct) or Glenn Gould’s Complete Collection 2015 full SBM Direct or basically a high proportion of Sony Music releases these days are still SBM Direct-based, if they merely downconvert the DSD masters they hold to 24/44.1 SBM Direct PCM. New recordings are rare, and that should explain something as well.
 
Last edited:
OP
Saidera

Saidera

Senior Member
Joined
Jul 18, 2020
Messages
388
Likes
309
Location
Adelaide, South Australia, Australia
SBM Direct testing

I wanted to see generally how SBM compares with Audacity shaped dither and then how SBM Direct compares with the two (however I realised that upon conversion to DSD via 65536 taps FIR, it will be difficult to compare perhaps because there is a DSD conversion stage - to overcome this I used two different converters).
The non-Sony 65536 taps software is explained at diyaudio:
https://www.diyaudio.com/forums/dig...stem-dsf-player-usb-ddc-dsd-amplifier-11.html

I initially used 24/44.1 derived from source file of 32 bit float, through SBM. But the 24/44.1 itself was force dithered by audacity. So the dither I ended up with was audacity+SBM. I was meant to see a pure SBM spectrum...

After intense struggles, I came up with a rudimentary test.
 
OP
Saidera

Saidera

Senior Member
Joined
Jul 18, 2020
Messages
388
Likes
309
Location
Adelaide, South Australia, Australia
Testing Method of dithers and noise shapers with a focus on DSD to PCM conversion (SBM Direct)

Although there are many DSD to PCM converters such as those in numerous playback software, editors, and names like Saracon, for now SBM Direct is the only one to be tested. SBM Direct has not been tested elsewhere, where Saracon and others were available. This is probably due to the rarity of Sonoma software.

To ensure compatibility, the source file of the 1 kHz test tone has to be in 24/44.1 PCM.

First, generate> chirp>sine starting from 1kHz and ending at 1kHz. Amplitude of the end of the ten seconds should be 0.000001. Then effect>amplify and reduce by 10 dB. Reduce too much and SBM will result in no sine wave at all. The area to analyse with the FFT is the last few seconds of the sine wave. All this is to fool the SBM into treating this as a music file. Then duplicate the mono Ctrl+D, then select the top track and make a stereo track from the two mono tracks. Again, this is because SBM needs stereo to work.

After this is done, simply convert to CD WAV via Audacity shaped dither, SEQUOIA, SBM and SBM Direct (via DSD Direct first, which apparently preserves the 20kHz of information), while SBM Direct will cut that noise, convert DSD to PCM, and then noise shape all in one stage at 64 bit. An alternative to DSD Direct may require lowering the gain by a lot to prevent errors during PCM to DSD conversion.

ssms.jpg
sequoia.jpg

1.PNG

2.PNG

This SSMS, much like Audiogate, uses Sonoma's Trim and Fade and gain adjustment multibit DSD editing. But no VST effects, because PCs were too weak back then.
3.PNG

4.PNG
 
OP
Saidera

Saidera

Senior Member
Joined
Jul 18, 2020
Messages
388
Likes
309
Location
Adelaide, South Australia, Australia
Results

Finally, compare FFTs of the last few seconds in Audacity and generally observe the noise curve shapes, how low the noise can go, and since I don’t understand dithers, perhaps learn how to judge which dithers to use?

original 24 bit sine.jpg

This is the original sine wave above.
audacity cd.jpg

Above is Audacity's shaped dither when 24 becomes 16 bit.
sequoia fft.jpg

This is SEQUOIA's standard triangular dither. Not so good?
powr2.jpg

This is SEQUOIA's pow-r #2 option.
SBM-1kHz.jpg

This above is Sony's old SBM from 24 to 16 bits.
sbmd 24 bit non dsd direct.jpg

This above is SBM Direct 24 bit non-Sony DSD Direct. From my understanding, 24 bit was converted to DSD, and then converted back to 24 bit, so there was virtually no added noise in the last few seconds. So instead, a much louder part was sampled here.
sbmd 24 bit sony dsd direct.jpg

This above is SBM Direct 24 bit via Sony DSD Direct. From my understanding, 24 bit was converted to DSD, and then converted back to 24 bit, so there was virtually no added noise in the last few seconds. So instead, a much louder part was sampled here.
sbmd 24 bit sony dsd direct 2.jpg

This above is SBM Direct 24 bit Sony DSD Direct taken again at the last few seconds. Does this prove anything? Is 24 bit SBM-D to be preferred, if so, what is the point since 24 bit has to be dithered to 16 bit to get the extra noise anyway. All we find from this is noise from DSD conversions.
sbmd 16 bit non dsd direct.jpg

I don't know what happened here. The amplitude of the DSD was much lower, yet noise was so high. This is non-DSD Direct derived.
sbmd 16 bit non dsd direct 2.jpg

I retook the FFT from the last seconds and this is 16 bit SBM Direct same as before.
sbmd 16 bit sony dsd direct.jpg

This is DSD from Sony's DSD Direct, converted to 16 bit wav. Looks better.
But see below for the hump rising from 14 kHz, same as the old SBM. Disappointingly enough.
sbmd 16 bit sony dsd direct 14 khz.jpg

After all that hype that this DSD CD format would be superior to plain PCM CDs, we end up with the same SBM curve, with the only difference being that it's sourced from DSD instead. 'Extended SBM' they used to call it. But then again, compared with other DSD to CD converters, isn't SBM Direct still superior? Many others probably use less than 32 k tap FIR filter. I prefer 65 k taps but I may have to write that software myself, as it only exists for PCM to DSD as linked to above.

Key Source files attached as zip files individually below.
 
OP
Saidera

Saidera

Senior Member
Joined
Jul 18, 2020
Messages
388
Likes
309
Location
Adelaide, South Australia, Australia
Key Source files
 

Attachments

  • Sony 1kHz.zip
    2.4 MB · Views: 92
  • 1kHz dff.zip
    2.7 MB · Views: 107
  • 16 bit SBMD Sony-1kHz.zip
    1.1 MB · Views: 76
  • 24 bit SBMD Sony-1kHz.zip
    1.4 MB · Views: 81
  • 16 bit SBMD 65k taps ver-1kHz.zip
    963.7 KB · Views: 71
  • 24 bit SBMD 65k taps ver-1kHz.zip
    1.3 MB · Views: 71
  • SBM-1kHz.zip
    1.2 MB · Views: 94

amirm

Founder/Admin
Staff Member
CFO (Chief Fun Officer)
Joined
Feb 13, 2016
Messages
44,376
Likes
234,529
Location
Seattle Area
Although there are many DSD to PCM converters such as those in numerous playback software, editors, and names like Saracon, for now SBM Direct is the only one to be tested. SBM Direct has not been tested elsewhere, where Saracon and others were available. This is probably due to the rarity of Sonoma software.
So what software you used for the conversions?
 
OP
Saidera

Saidera

Senior Member
Joined
Jul 18, 2020
Messages
388
Likes
309
Location
Adelaide, South Australia, Australia
With the utmost gratitude to Blumlein 88 and Pluto for their kind and patient assistance in this utterly pointless exercise!

Blumlein 88 introduced me to the following:

https://www.gammaelectronics.xyz/s_1993_01_sony-sbm.html

This article shows SBM's noise curve vs one from Stanley Lipschitz.

index.php



Blumlein 88 confirmed that you can turn off all dither in Audacity when it saves a file. So whatever was in the 44/24 file is all that would be saved.

Saidera: And when I selected none as dither it deleted the sine wave. And if I selected shaped it appears.
So I asked, how come CD WAV contains -120db info in FFT?

Blumlein 88: Yes if you save a -120 db signal in non-dithered 16 bit you lose it. The lowest significant bit shows -96 db, and anything below that is lost. That is also the benefit of dither. By adding low level noise you can encode below the LSB and recover the signal. Some dithers can record and retrieve nearly -130 db around 4 khz in a 16 bit format.

How can that work? The dither means at no time is the total signal below the LSB. So with the noise spread over the spectrum properly that lowest bit can encode a noisy version of an even lower signal because the LSB doesn't ever "turn off". If you amplify such a signal enough you'll get the tones back, but there will be a little hiss noise mixed in with it at a low level.

Saidera: Too complex... but intriguing! Cooler than compression codecs. SBM is way above the 138 db of audacity. And no 4 khz dip either. When it's displayed linearly, you will see noise rising from 14 kHz onwards, just like EVERY SBM FFT floating around the net and Heddle et al in that diagram! OK, SBM fawning can be put aside.

Blumlein 88: I'll take some liberties with details, but think I can give you a simple idea how dither works.

Imagine I have a 1 khz tone at -103 db fed to a non-dithered 16 bit ADC. It will record nothing because the signal is lower than -96 db of the LSB (least significant bit). Now if I have a 1 khz tone at -103 db and a 3 khz tone at -103 db the total level of those combined tones is -93 db, and fed into the same ADC they will get recorded. An FFT analysis will show both tones and show them at -103 db, but at no time was the total signal to the ADC less than -93 db.

Dither does the same thing. Dither is noise with small low level signals at every frequency up to 20 khz. I can put some higher levels at higher frequencies and leave some out or have them at a much lower level where hearing is most sensitive. So you have these thousands of very low level tones and leave room for signals well below -96 db in the 3-5 khz region. But at all times the total signal for all of this is above -96 db. Typically dither will keep the total noise level to at least -90 db. Some even utilize two bits and have the total dither signal around -84 db. But with proper shaping you can record and retrieve signals well below -96 db in some frequency bands.

Now again I skirted some technical details, but it gives you the flavor of how it works.

https://deltaw.org/

If you've not used Deltawave, it is very useful for comparing various digital files. It started out pretty simple, but has grown to be capable of all sorts of things. It is basically still a null testing software. Put two versions of music in, subtract one from the other and see what differences remain.
Paul Kane who is a member here (pkane) wrote it. You might want to check it out.

Saidera: Thanks! I will check it out.
 
OP
Saidera

Saidera

Senior Member
Joined
Jul 18, 2020
Messages
388
Likes
309
Location
Adelaide, South Australia, Australia
So what software you used for the conversions?

For converting PCM to DSD I used:
Wave to Dsdiff Converter - Version1.5 - GUI
from http://wpup.html.xdomain.jp/
with PCM-DSDコンバータ version1.5 from http://ama-audio.seesaa.net/article/287217665.html
as well as Sony VAIO DSD Direct http://www.vaio.sony.co.jp/vaio/solution/DSDdirect/feat2.html
https://www.sony.com.au/electronics/support/articles/S500024364
For converting DSD to PCM I used:
SonicStage Mastering Studio 2.4.03 Recorder Edition (bundled with PCM-D50)
http://www.vaio.sony.co.jp/vaio/solution/SonicStageMasteringStudio/

Thank you for your interest in this. I am surprised and very humbled.
 
OP
Saidera

Saidera

Senior Member
Joined
Jul 18, 2020
Messages
388
Likes
309
Location
Adelaide, South Australia, Australia
The next investigation on SBMD is a comparison of Yo-Yo Ma's DSD recording on SACD with the SBMD CD of 2009 from his box set. So the first file will be the SBMD CD, and the second shall be downconverted DSD via my own SBMD to 24 bit and 16 bit. The hypothesis is that the SBMD CD should be identical to the 16 bit file, and the 24 bit file should be of better quality. The SACD must be a pure DSD recording but it may be a conversion from non-DSD.
 
OP
Saidera

Saidera

Senior Member
Joined
Jul 18, 2020
Messages
388
Likes
309
Location
Adelaide, South Australia, Australia
Dithering is usually performed to improve the quantization noise when dropping a high-resolution sound source with a bit depth of 24 bits or more to CD format (16 bits). Should dithering processing be performed for the following? When writing to CD format from a high-resolution recording and for exporting a high-resolution sound source to AAC (MP3).

The principle of dithering is technical and difficult to understand, so I used a 1khz sine wave (24bit high-resolution sound source) and converted it to CD format (16bit) to see how it would behave in the spectrum if dithering was added.

In the case of triangle dither addition, noise was added evenly over the entire bandwidth, and if we focus only on harmonic distortion, the improvement was better than without dither. (THD: 0.0032% → 0.00174%). In the case of noise shaving with dither: Although the high frequency noise was larger, the noise in the 1khz to 5khz band, where audible sensitivity is higher, was reduced more than that without dither.

In recordings from vinyl records, the noise floor is around -100db, and I don't think it makes sense to add more dither noise. Also, the spectrum of the triangle dithered signal has a flatter noise floor in the high frequency range (around 19khz and above) than that of the undithered signal, but this is due to the fact that dither noise is added to the entire bandwidth.

The sound files distributed by Apple Music are delivered in the dither-free AAC format using Apple Digital Masters technology. The gist of this is that dithering is not necessary because the AAC encoder supports 24-bit. The noise floor is at about the same level as the original 24-bit high-resolution file, and is much smaller than the 16-bit quantization noise of a CD. If the bit depth is 16-bit, dithering is necessary, and if it is 24-bit, dithering is unnecessary.
 
OP
Saidera

Saidera

Senior Member
Joined
Jul 18, 2020
Messages
388
Likes
309
Location
Adelaide, South Australia, Australia
The next investigation on SBMD is a comparison of Yo-Yo Ma, Kathryn Stott - La Belle Époque (2003) track 4. Violin Sonata No.1 in A Major for Violin and Piano, Op.13 III. Allegro vivo.
Most likely recorded in DSD due to the CD liner notes explaining that it uses DSD and SBMD. The SACD is downconverted via SBMD to 24 bit and 16 bit. This is compared with the SBMD CD of 2009 from his box set.
24b sbmd violin.jpg

Above is 24 bit SBMD
new sbmd violin.jpg

Above is 16 bit SBMD
original sbmd violin.jpg

Above is the 2009 CD Box 16 bit SBMD
new 24b 2 2 secs sbmd violin.jpg

Above is 24 bit SBMD
new 2 sbmd violin.jpg

Above is 16 bit SBMD
original 2 sbmd violin.jpg

Above is the 2009 CD Box 16 bit SBMD
 
OP
Saidera

Saidera

Senior Member
Joined
Jul 18, 2020
Messages
388
Likes
309
Location
Adelaide, South Australia, Australia
There are slight differences between my CD SBMD and that of Sonoma from the 2009 remaster CDs, which could be attributed to selection error for the FFT - the first 2 seconds of audio was used.
 
OP
Saidera

Saidera

Senior Member
Joined
Jul 18, 2020
Messages
388
Likes
309
Location
Adelaide, South Australia, Australia
The next investigation on SBMD is a comparison of Yo-Yo Ma - Plays The Music Of John Williams (2002) track 7. Three Pieces For Solo Cello - Pickin'
Most likely recorded in DSD due to the CD liner notes explaining that it uses DSD and SBMD. The SACD is downconverted via SBMD to 24 bit and 16 bit. This is compared with the SBMD CD of 2009 from his box set.

24 bit williams 1.jpg

Above is 24 bit SBMD
16 bit williams 1.jpg

Above is 16 bit SBMD
16 bit williams 1 original.jpg

Above is the 2009 CD Box 16 bit SBMD
24 bit williams 2.jpg

Above is 24 bit SBMD
16 bit williams 2.jpg

Above is 16 bit SBMD
16 bit williams 2 original.jpg

Above is the 2009 CD Box 16 bit SBMD
 
OP
Saidera

Saidera

Senior Member
Joined
Jul 18, 2020
Messages
388
Likes
309
Location
Adelaide, South Australia, Australia
There are significant differences between my SBMD files created from SACD and that of Sonoma from the 2009 remaster CDs, which could be attributed to the 2009 master DSD source file or PCM mastering involving a gain adjustment of 4 dB or more. Selection error for the FFT in selecting the first 2 seconds of audio has been minimised, however it was unnecessary.
 

Blumlein 88

Grand Contributor
Forum Donor
Joined
Feb 23, 2016
Messages
20,524
Likes
37,057
Use a larger FFT size. Especially with the wider bandwidth. Audacity can do up to 65,536. (64K FFT).
 
OP
Saidera

Saidera

Senior Member
Joined
Jul 18, 2020
Messages
388
Likes
309
Location
Adelaide, South Australia, Australia
I also attempt an ABX by using a 24 bit wav (closer to DSD original?) vs 16 bit SBMD (characteristic of PCM maybe?) using both tracks but not using the 2009 CD files (due to loudness differences), instead using the 16 bit files I created.

foo_abx 2.0.6d report
foobar2000 v1.5.4
2021-07-11 16:57:35

File A: 01-Faur・ Sonata No. 1 24b.wav
SHA1: 08d7494d7212c1f220e06b8d94c8333c2498cfcc
File B: 01-Faur・ Sonata No. 1 in A Major for Violin and Piano, Op. 13 - I~.wav
SHA1: d564b7a89238934cb7cb8b2cd990da7134704620

Output:
ASIO : SONATA-HD
Crossfading: NO

16:57:35 : Test started.
17:02:31 : 01/01
17:02:54 : 01/02
17:03:15 : 02/03
17:03:39 : 03/04
17:03:56 : 03/05
17:05:03 : 03/06
17:05:32 : 04/07
17:05:59 : 05/08
17:06:27 : 05/09
17:06:43 : 05/10
17:07:55 : 06/11
17:09:34 : 06/12
17:10:10 : 07/13
17:11:21 : 08/14
17:11:46 : 09/15
17:13:30 : 09/16
17:13:30 : Test finished.

----------
Total: 9/16
p-value: 0.4018 (40.18%)

-- signature --
659c7550e59c16a9b1d05c924fcb792bb68cb5ed

foo_abx 2.0.6d report
foobar2000 v1.5.4
2021-07-11 17:15:07

File A: 02-Three Pieces For Solo Cello - Pickin'.wav
SHA1: ed6ea9d63eb38cd70a1647020f443598b11d93d9
File B: 02-Three Pieces For Solo Cello 24b.wav
SHA1: fe9ec4f858c1daa2f87d4ab3b359efacef4a18d0

Output:
ASIO : SONATA-HD
Crossfading: NO

17:15:07 : Test started.
17:19:37 : 00/01
17:21:06 : 01/02
17:21:21 : 02/03
17:21:52 : 02/04
17:22:01 : 02/05
17:22:25 : 02/06
17:22:32 : 03/07
17:22:39 : 03/08
17:23:09 : 03/09
17:23:33 : 03/10
17:24:09 : 04/11
17:24:53 : 04/12
17:25:02 : 05/13
17:25:45 : 06/14
17:26:03 : 07/15
17:26:56 : 07/16
17:26:56 : Test finished.

----------
Total: 7/16
p-value: 0.7728 (77.28%)

-- signature --
687ee81dd8b7095d73120bbb90b1db7c134569ea
 
OP
Saidera

Saidera

Senior Member
Joined
Jul 18, 2020
Messages
388
Likes
309
Location
Adelaide, South Australia, Australia
Since Audacity can do up to 65,536 FFT size, here are the new FFTs for track 7. Three Pieces For Solo Cello - Pickin'
new wav 1.jpg

Above is my own 16 bit SBMD
original wav 1.jpg

Above is the 2009 SBMD DSD CD
new wav 2.jpg

Above is my own SBMD
original wav 2.jpg

Above is the original CD

If there is any noise shaping, it does not appear obvious to me. The 2009 CD tapers off at the end unlike my SBMD for 16 bit. Immediately one questions whether the SBMD I have is faulty or watered-down by VAIO software crew and Nishio et al. But the loudness difference between the 2009 CD and my SBMD is no longer so obvious. There could be errors in generating the FFT - this time I used 1.500 seconds but it is never an exact 1.5, it is always off the mark somewhat.

Luckily, the violin sonata I did first suggests that the VAIO SBMD I have is probably working just like the SBMD in Sonoma software, and is not watered down.

But the excruciating and challenging ABX would likely suggest to anyone to choose 16 bit SBMD if given a choice between a DSD CD and 24/44.1 Hires download from SBMDed DSD, unless one really wants to use a non-noise shaped file.
 
Top Bottom